fix(data): use Energon validation recipe for val split

[caohch-1]

What does this PR do ?

Use Energon's validation dataset recipe for the validation split instead of constructing validation data through the training recipe.

Changelog

• Route the training split through get_train_dataset() and the validation split through get_val_dataset().
• Keep training-only shuffle configuration out of the validation path while preserving existing dataset-loading options.
• Add regression coverage for validation dataset recipe selection and update existing validation dataloader tests to use the validation builder.

Validation

• A focused regression harness confirms that the pre-fix implementation selects get_train_dataset() for validation and the updated implementation selects get_val_dataset().
• The same harness confirms that the training split continues to use get_train_dataset().
• git diff --check passes.
• Python compilation checks pass for both changed files.

GitHub Actions CI

This is an external contribution, so GitHub Actions CI will require approval from an NVIDIA developer.

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you add or update any necessary documentation? No documentation changes are needed for this fix.
• Does the PR affect components that are optional to install? No.
  • Reviewer: Does the PR have correct import guards for all optional libraries?

Additional Information

Fixes #5626

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[yaoyu-33]

validation_task_encoder will this always exist? or required

[caohch-1]

Yes. validation_task_encoder is initialized in __init__ as validation_task_encoder or self.task_encoder, so it falls back to the training task encoder when a separate validation encoder is not provided. The existing val path already used self.validation_task_encoder; this PR only changes the dataset builder from get_train_dataset() to get_val_dataset().

If both encoders are None in direct construction, the value can still be None, but that is pre-existing behavior and is not introduced by this change.