Remove PyTorch Lightning dependency

[MilesCranmer]

• Removes the PyTorch Lightning dependency which was too heavyweight
  • (Ideally we could also remove the Torch dependency, but seems like a bit more work)
• Switches from pickle to safetensors which is a faster and safer way to store model weights
  • Also I upload the file used to do this translation, for posterity (in case had fine-tuned a custom model)
• Stores model hyperparameters in accompanying json files

Also I made the README example of deep regressor a bit better.

[dtamayo]

Awesome...we were also running into trouble with pickled model files with Caleb. Pytorch seemed to suggest this to avoid pickling:

https://pytorch.org/tutorials/beginner/saving_loading_models.html#saving-loading-model-for-inference

Safetensors is more robust?

[MilesCranmer]

Yeah safetensors is pretty much all that they recommend there, plus additional stuff to make it even safer (since pickle allows arbitrary code execution).

Also don't merge this yet; trying to fix some issues...

[dtamayo]

If you figure it out, could you tell Caleb how you did it so we can save his models your way? With pickle it also causes problems being loaded in a different module from the one in which it was saved.

[MilesCranmer]

What is his branch?

[dtamayo]

working on the caleb branch

[MilesCranmer]

Thanks!

Okay I think this PR is ready for review. I also removed the pandas and matplotlib dependency for good measure.

[MilesCranmer]

Re: Caleb's branch, I don't see data for the model stored anywhere (?)

Also is there a test to try it out on for debugging?

[dtamayo]

Thanks so much Miles. Will merge these changes into the caleb branch and add a test