Fix: apply sequence packing loss wrapper in SFT validation

[XChen-Zero]

What does this PR do?

This PR fixes a bug in SFT validation when use_sequence_packing is enabled.

In the current implementation, train_step wraps the loss function with
SequencePackingSFTLossWrapper, while val_step directly uses the raw
loss function. This causes validation to fail when sequence packing is enabled.

This PR applies the same loss wrapper logic in val_step to keep training and
validation behavior consistent.

Why is this needed?

When use_sequence_packing is enabled:

• Training uses SequencePackingSFTLossWrapper to correctly handle packed sequences
• Validation does not apply the wrapper
• As a result, forward_step receives incompatible inputs and validation fails

This leads to runtime errors during validation and makes SFT training with
sequence packing unusable.

Changes

• Apply SequencePackingSFTLossWrapper in val_step when use_sequence_packing is enabled
• Align loss handling between train_step and val_step

Backward Compatibility

• No behavior change when use_sequence_packing is disabled
• Fully backward compatible

Related Issue

N/A

[CLAassistant]

All committers have signed the CLA.