Upgrade to pytorch-lightning 2

[AmirAlavi]

This migrates the code to the latest major version of pytorch-lightning. I tried to do it with minimal changes to stay true to the original design philosophy of Casanovo. Feedback is much appreciated.

Summary of changes (I followed this guide to upgrade):

• LightningLite is deprecated
  • Switch to Fabric
• auto_select_gpus parameter for Trainer is deprecated
  • Use find_usable_cuda_devices in _get_devices instead. Also, devices can't be 0, so now _get_devices falls back to auto
• Relying on passed in outputs in on_predict_epoch_end is deprecated
  • Use self.trainer.predict_loop.predictions instead
• Passing a dictionary to LightingModule.log is deprecated
  • Log as scalar with the "mode" (e.g. "train" or "valid") embedded in string key to the same effect
• strategy parameter for Trainer cannot be None
  • Return "auto" if no cuda devices

[wfondrie]

Hi @AmirAlavi - thank you for your PR, but we've actually already begun this migration and it will be present in the next release. You can check out the changes in #176.

Casanovo uses a Git Flow-type of system, where our in-development work goes into the dev branch before being migrated into main. There's a lot of other updates waiting to be merged into main too!

[AmirAlavi]

Ah got it, thanks @wfondrie, don't know how I missed that PR. Thanks! Will stay tuned.

[wfondrie]

No worries - we need to update our contribution docs to be more accurate since the change. Sorry for the confusion!