fix(datasets): align reasoning + assistant loss masks for left padding

[khazic]

What does this PR do?

Fix loss-mask misalignment for left-padding tokenizers in
format_chat_template. Affects both the assistant mask and the
reasoning mask paths.

Background

_build_multiturn_assistant_mask and _build_reasoning_mask compute
their span indices from unpadded (padding=False) tokenizations
because they walk the formatted text to find role boundaries. The mask
arrays they return are sized to the padded input_ids, however --
and when tokenizer.padding_side == "left" (e.g. Phi-3,
some Qwen-VL configs) the actual content is right-aligned in
input_ids. The mask positions then fall in the left-padding region,
the subsequent attention_mask-based zeroing wipes them, and training
silently learns from an all-zero loss mask for ~80% of samples.

Relationship to #2312

#2312 fixes this for the assistant mask path with an inline shift.
This PR is a strict superset of that fix:

1. Extracts the shift into a reusable _maybe_shift_mask_for_left_padding
   helper.
2. Applies it to both the assistant mask path (same site as feat(speculative): add Phi-3 support for EAGLE-1/2/3 + correctness fixes #2312)
   and the reasoning mask path (which feat(speculative): add Phi-3 support for EAGLE-1/2/3 + correctness fixes #2312 missed).
3. Adds unit tests for the helper (8 cases covering happy / no-op /
   edge paths).

#2312 has since landed; the rebase conflict at the assistant-mask site has
been resolved by replacing the inline shift with the helper call.

Changelog

• components/datasets/llm/formatting_utils.py:
  • Add _maybe_shift_mask_for_left_padding(mask, tokenizer, attention_mask)
    helper. Three short-circuit guards make it a no-op for
    right-padding tokenizers, missing attention_mask, and
    pad_len == 0.
  • Replace the inline shift from feat(speculative): add Phi-3 support for EAGLE-1/2/3 + correctness fixes #2312 with a call to the helper after
    _build_multiturn_assistant_mask.
  • Call it after _build_reasoning_mask (the bug feat(speculative): add Phi-3 support for EAGLE-1/2/3 + correctness fixes #2312 missed).
• tests/unit_tests/datasets/llm/test_shift_mask_left_padding.py:
  8 unit tests using a SimpleNamespace fake tokenizer (no HF
  dependency), covering right-padding no-op, left-padding shift,
  zero pad_len no-op, attention_mask=None, missing
  padding_side attribute (default right), all-padding,
  single-content-token.

Verification

Unit tests pass locally:

pytest tests/unit_tests/datasets/llm/test_shift_mask_left_padding.py -v

For right-padding tokenizers (Llama / Qwen / Mistral / ...) the helper
returns the original list object unchanged -- no behavioural change.

For left-padding tokenizers (Phi-3, ...) the previously-silent
all-zero loss_mask is now correctly aligned with content positions.

Before your PR is "Ready for review"

Pre checks:

• Contributor guidelines followed
• Unit tests added (test_shift_mask_left_padding.py)
• No documentation changes needed -- internal helper.

[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.

[HuiyingLi]

/ok to test 861fa9d

[HuiyingLi]

/claude review

[HuiyingLi]

/claude review

[HuiyingLi]

/ok to test f35f088