Fix: Allow the user to select whether to drop the last sentence

[Rive-001]

This PR adds a configurable drop_last parameter to the PerturbationDataModule class, providing users with control over whether incomplete sentence sets should be dropped during data loading.

Changes

New parameter: Added drop_last: bool = False parameter to PerturbationDataModule.__init__()
Documentation: Updated docstring to describe the new parameter's behavior
Implementation: The parameter now controls the drop_last behavior in PerturbationBatchSampler instead of being hardcoded to False

Behavior

When drop_last=True, sentence sets that are smaller than cell_sentence_len will be dropped during data loading.

The default value remains False to preserve backward compatibility, meaning incomplete sentence sets will be kept by default.

This PR passes all existing unit tests

[abhinadduri]

to avoid dropping the same cell sentence each time, can we randomize the order of the mini batch each epoch?

[Rive-001]

to avoid dropping the same cell sentence each time, can we randomize the order of the mini batch each epoch?

As I understand, after merging #50 , we're creating new batches every epoch. So the contents of the last partial batch would change every epoch. Therefore I don't think we're dropping the same set of sentences every epoch. But to make sure, I could add a unit test to check this.