`BatchSizeFinder` with method `fit` and separate batch_size attributes for train and val (e.g., `self.train_batch_size` and `self.val_batch_size`)

[ibro45]

Outline & Motivation

I suggest allowing batch_arg_name to accept a list of arg names. E.g. `tuner.scale_batch_size(..., batch_arg_name=["train_batch_size", "val_batch_size").

Pitch

Fit uses both train and val dataloaders. They can have their own batch sizes.

Additional context

No response

cc @lantiga @justusschock