Implement DPM Solver Multistep Scheduler

[mspronesti]

Hi, this PR aims at integrating the DPM Solver Multistep Scheduler in this Rust version, as requested in #17 .

A couple of disclaimers:

• the implementation contains all the features of the Python version, but the optional thresholding, described in this paper. I plan to add it in a future PR, if you agree.
• I don't like the initialization of model_outputs. Unfortunately tch::Tensor doesn't implement the Copy trait, so I couldn't find a leaner way. Feel free to leave a suggestion if you have a better idea!

• differently from the other two schedulers, the step method requires self to be mutable as I need to update the attribute model_outputs. Therefore, to run the examples, make sure to create a mut scheduler:

  let mut scheduler = DPMSolverMultistepScheduler::new(n_steps, Default::default());

  and to call to_owned when iterating over the timesteps (e.g. here):

  for (timestep_index, &timestep) in scheduler.timesteps().to_owned().iter().enumerate() 

Cheers :smile: .

[mspronesti]

I just noticed I missed your first question. I "tested" the implementation using the snippet I shared with you in my last PR (changing the scheduler, clearly :)). The codebase should be almost 1:1 with the python version (or, at least, that's what I tried to do).

[LaurentMazare]

Good, could you maybe add a few permalink to the python version in the scheduler code? This is nice for people that would want to see how the code was converted.

[mspronesti]

Sure, is there a specific part you suggest to "annotate" with a permalink ? For instance, here I could replace this comment with a permalink to the python version :) The rest of the code seems (to me) fairly similar to the python implementation, including the comments and the name of the variables.

[mspronesti]

Not sure why one of the checks failed. Seems the classic error one gets when .../cargo/index is corrupted and just needs to be deleted and re-created, or perhaps a network error. Maybe re-running the action solves it ?

[LaurentMazare]

Rerunning fixed indeed this test, I've merged the changes, thanks for the PR!