Trainer Usability Improvements + Lightning Compatibility

[dilyabareeva]

• For training-based metrics and benchmarks, the Trainer object can either be a Lightning trainer or inherit from quanda BaseTrainer.
• Quanda now only includes a trainer implementation with simple arguments like learning rate, optimizer, scheduler, etc. So the user options are: 1) create their own Trainer inheriting from BaseTrainer, 2) use our default Trainer implementation, 3) Lightning trainer in combination with a Lightning module passed instead of a torch.nn.Module
• Added tests to check if quanda works as expected with Lightning modules inserted in place of torch.nn.Modules - useful for many users.
• Added an optional init_model argument for DatasetCleaning - the user might want to train a model from scratch.
• Made sure that when we train models we do this on a copy of a module to avoid unexpected side effects.

Closes #93.

[gumityolcu]

Hi Dilya, I left a couple of minor comments. I really like the simpler training strategies. Much easier to understand and use (and document :))