[vla] fix: Add batch drop-last configuration

[XueSongTap]

Summary

This PR makes batch_drop_last easier to configure for VLA training.

The option controls whether the final incomplete batch is dropped and is applied consistently to both the sampler and the DataLoader.

Configuration

Default behavior

When no option is specified, the final incomplete batch is preserved:

Enable drop-last

Add the following flag to the training command:

--batch-drop-last

When enabled, the final batch is dropped if it contains fewer samples than the configured batch size.

Explicitly disable drop-last

The behavior can also be explicitly disabled with:

--no-batch-drop-last

Behavior

Configuration	batch_drop_last	Behavior
Option omitted	False	Preserve the final incomplete batch
--batch-drop-last	True	Drop the final incomplete batch
--no-batch-drop-last	False	Preserve the final incomplete batch

This configuration is applied to:

• The distributed sampler
• The DataLoader

For distributed training, enable --batch-drop-last when every rank should process the same number of full-sized batches.

Example Updates

The PI0.5 DDP, ZeRO-1, and FSDP fine-tuning examples now explicitly enable:

--batch-drop-last