refactor(data): move VLM sequence batching to collate

[yaoyu-33]

Stacked on #4307. Current order: #4169 -> #4307 -> #4315.

Implements issue #4041 Step 2 by moving generic VLM padding, truncation, and in-batch packing preparation from vlm_step into collate-time helpers. The VLM training step now consumes already-collated sequence tensors and packed metadata.

Validation status:

• Targeted static checks passed locally: py_compile, ruff check, ruff format --check, and git diff --check.
• Full pre-commit is currently blocked on this workstation by the known nvidia-resiliency-ext wheel/platform resolver issue.
• Internal loss-parity validation against feat(data): add text chat collate for unified HF datasets #4307 head is pending; this PR remains draft until that completes.

[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.