Add options for eval and gradient required #78 - rebase

[jatkinson1000]

This is an updated version of #78 rebased onto main after the GPU changes by @jwallwork23

@ElliottKasoar's comment on the original PR:

Resolves #73

Adds flags in all(?) functions that operate on tensors (tensor creation, model loading, forward) to optionally disable autograd, which should improve performance for inference.

Also adds a similar flag to set evaluation mode for the loaded model.

• Evaluation mode

  • Contrary to my initial comments in #73, from testing evaluation mode does appear to be preserved, both between saving and loading TorchScript, and when applied to the loaded model.
  • In most cases evaluation mode is therefore likely to already be set, but I think it's useful to have the option to change it, particularly if FTorch may be extended to facilitate training (#22).

• NoGradMode

  • Enabling or disabling gradients is more complicated, as it defined via a context manager, which only appears to define the behaviour within its own scope, and so it seems necessary to enable/disable gradients before every code block that operates on tensors (similar to the Python equivalent with torch.no_grad():).

• InferenceMode

  • No changes are currently included, but it would be good to support InferenceMode too eventually, as it should provide further performance benefits over NoGradMode.
  • However, it has stricter requirements, and the mode was only added (as a beta) in PyTorch 1.9, so we would need to be careful if we want to support older versions.

• Model freezing

  • No changes are currently included, and less directly applicable to the main FTorch library, although there are still interactions e.g. freezing the model can allow InferenceMode to be enabled when loading the model.
  • Freezing is currently the "default" when tracing in pt2ts.py, but not for scripting, despite potentially improving performance.
  • Freezing appears to (sometimes) introduce numerical errors when saving the reloading (differences ~10^-6), and can seem to lead to issues loading with Forpy too.

(For more general explanation of autograd/evaluation mode, see autograd mechanics).

Note: I've also removed the old, commented out torch_from_blob function.

[jatkinson1000]

Hi @ElliottKasoar after the work @jwallwork23 did there were some conflicts with your PR in #78.

I have done my best to rebase your work, but please could you take a look and see if everything seems in order to you?

@TomMelt You originally reviewed this PR and approved, but if you could take a quick glance and re-review it would be appreciated - since @jwallwork23 restructured the order of functions in the files and added additional arguments in the same place as @ElliottKasoar some of the merge conflicts got a little hairy so I may have missed the odd thing!

[jatkinson1000]

Before we merge we need to:

• [x] Add some docs for these new options
• [ ] Add an example showing the use of these new options? Or perhaps not given we are not training yet and just use the defaults...?

[jatkinson1000]

Added a note to the FAQ about eval and no_grad settings. A detailed example will perhaps wait until these are used as part of #111 since for now they are the sensible defaults for running inference.

I will also move some of @ElliottKasoar's points in his original comment to separate issues for future consideration.

Squashing and merging shortly.