Use one Optimizer per feedback layer

[lebrice]

Uses one optimizer per feedback layer in the PL implementation of DTP.

ParallelDTP uses a single optimizer for all feedback layers.

@amoudgl Let me know what you think.

Also: The ordering of the hyper-parameters for the feedback network (lr, iterations, noise scales, etc) in my implementation are a bit annoying to work with, now that I look at it.

• It would probably be a lot simpler to just have self.F and self.G (where x_r = self.G[i](self.F[i](x)).
• The only real downside of doing it this way is just that self.G wouldn't be usable end-to-end like a regular nn.Sequential, (i.e. self.G(self.F(x)) through the entire network wouldn't work)

[amoudgl]

@lebrice I have taken a pass and added a few comments. Once these comments are resolved, we'll have a bug-free parallel DTP implementation that Sean can use for hyperparam tuning.

I'll update the legacy tests and push changes in this branch soon.

[amoudgl]

Also: The ordering of the hyper-parameters for the feedback network (lr, iterations, noise scales, etc) in my implementation are a bit annoying to work with, now that I look at it.

I agree but I think we should keep it this way for now because it's just easier to test/validate + save models where you can just use existing codebases structure off the shelf.

[amoudgl]

@lebrice I updated Maxence legacy implementation to use a separate feedback optimizer per layer in c566066 and updated tests accordingly -- all legacy unit tests now pass.