119 improve api

[maho3]

This PR implements numerous structural changes to the configurations of the Data/Inference/Validation stages, without much changes to how they work on the backend. The general theme of this PR is to add various quality-of-life improvements that make it easier to understand the configurations and use ltu-ili for distributed testing. Here's a summary of what has changed.

Removed unnecessary configurations:

• Removed the need to specify backend in the ValidationRunner. This is now gathered automatically from the imports.
• Removed the need to specify n_walkers in DelfiRunner (inference stage). It didn't make sense to specify n_walkers before the sampling stage, and it is redefined anyway in the EmceeSampler during validation.
• Removed the need to specify n_data and n_params in the DelfiRunner configuration. This info is now gathered automatically from the provided dataloader.
• Removed the n_data configuration of the FCN network. It was simply the dimensionality of the output of the network. It can now be configured as the last entry in n_hidden
• Removed the inference_class argument of the inference Runner objects and replaced it with engine. Previously, users had to load an engine object from file before passing it into the Inference runners (e.g. inference_class = sbi.inference.SNPE. Now users only need to specify a string (e.g. engine='NPE' or NLE/NRE/SNPE etc) and we figure out what object inference_class dependent on the backend.
• All path-like arguments (e.g. out_dir, in_dir, etc) can be specified with str inputs instead of pathlib.Path. We convert them all to Path's on the backend.

Renamed things to make them more obvious:

• Changed output_path -> out_dir and posterior_path -> posterior_file to match the naming conventions in loaders.py. Now, the ValidationRunner will only load the posterior from out_dir / posterior_path and subsequently save metrics in out_dir. I didn't see a reason why one would want to save metrics in a different directory from their posterior.
• Similar to the last change, output_path -> out_dir in the Inference stages.
• All stages can now be loaded from their top-level category. E.g.

  from ili.dataloaders import StaticNumpyLoader
  from ili.inference import SBIRunner
  from ili.validation import ValidationRunner

  Note, The old imports (e.g. from ili.inference.runner_sbi import SBIRunner still work for backwards compatibility.

Conformed the multiple Inference backends to a universal configuration:

• Added InferenceRunner in ili/inference/runner.py to act as a universal engine for the Inference stage. It's configuration takes backend and engine as parameters, and then uses these to go off and find the write object for your inference. For example, backend='sbi' and engine='NLE' will load the configuration in SBIRunner, whereas backend='sbi' and engine='SNLE' will load it in SBIRunnerSequential. This new configuration should be preferred, but it does not affect the way we previously specified SBIRunner/SBIRunnerSequential/DelfiRunner etc, so it should be backwards compatible. Adds appropriate tests.
• Added ili/utils/ndes_pt.py and ili/utils/ndes_tf.py which provide the load_nde_sbi and load_nde_pydelfi which provide utilities for loading neural architectures from given configurations for each backend. This unifies the configuration with which we specify neural architectures between sbi and pydelfi, making things considerably simpler. These functions were used to refactor SBIRunner and DelfiWrapper. Adds appropriate tests.
• Added the ability to overload a loaded config file in each from_config function of the Data/Inference/Validation stages. This can be very useful if you want to iterate (in python) over many small changes to a configuration file. For example, one could change the training rate and the out_dir to save many models to different directories, each of which trained with a different training rate. Addresses #119
• DelfiWrapper's and EmceeSampler's sampler functions were fixed to only produce nsteps samples, matching the behavior pyro samplers. Previously, they were producing nsteps*num_chains samples.
• DelfiRunner now saves a summary of train/validation loss, following SBIRunner.

Miscellaneous:

• Updated the jupyter notebook example with the new interface.
• Updated CONFIG.md accordingly.
• Added _BaseRunner in ili/inference/base.py as a template for sbi/pydelfi runners

Note a lot of files have changed, because we needed to propagate the naming changes to all the examples and tests. I'm happy to walk the reviewer through the listed changes.

[codecov-commenter]

Codecov Report

Attention: 10 lines in your changes are missing coverage. Please review.

Comparison is base (a11f4bb) 97.30% compared to head (3dbc6f6) 97.27%.

Files	Patch %	Lines
ili/dataloaders/loaders.py	79.31%	6 Missing :warning:
ili/inference/runner_sbi.py	96.00%	2 Missing :warning:
ili/inference/runner_pydelfi.py	97.29%	1 Missing :warning:
ili/utils/ndes_pt.py	95.45%	1 Missing :warning:

:exclamation: Your organization needs to install the Codecov GitHub app to enable full functionality.

Additional details and impacted files

```diff @@ Coverage Diff @@ ## main #123 +/- ## ========================================== - Coverage 97.30% 97.27% -0.03% ========================================== Files 18 22 +4 Lines 1113 1249 +136 ========================================== + Hits 1083 1215 +132 - Misses 30 34 +4 ```

:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.

[CompiledAtBirth]

Currently running tests on a duplicate of this branch (seq_test to use a more recent torch than the setup.cfg).

[DeaglanBartlett]

@CompiledAtBirth how did the tests do on the duplicate branch?

[DeaglanBartlett]

From @CompiledAtBirth on slack:

For the SBISimulator loader specifically, the simulator function has to output a torch.Tensor of torch.Size((1,n)). The case is easy to handle once the user knows his simulator function must return that.

To avoid this mistake, please can you make sure it is well documented in the code and in the tutorial? We might as well do it in this PR as it is a minor thing to add

[maho3]

From @CompiledAtBirth on slack:

For the SBISimulator loader specifically, the simulator function has to output a torch.Tensor of torch.Size((1,n)). The case is easy to handle once the user knows his simulator function must return that.

To avoid this mistake, please can you make sure it is well documented in the code and in the tutorial? We might as well do it in this PR as it is a minor thing to add

Added comment in the code and tutorial.

[maho3]

The numpy seed is updated in PlotSinglePosterior if x_obs is None and a seed is provided

This is rather difficult to test because we don't return what x_obs we choose, and the used np seed is difficult to access. I think this test isn't necessary, given the code is pretty clear as to what its doing.

[maho3]

Check ValidationRunner.load_posterior_sbi() works even if the posterior does not have a name attribute

I just removed this functionality of load_posterior_sbi, because there should never be a case wherein an sbi posterior doesn't have a name. It is given "" by default.