fix(model): use dynamic temporal video token counts

[cuichenx]

What does this PR do ?

Adds an explicit, processor-driven temporal-video resize mode for Nemotron Omni Energon training so each tubelet receives the number of image placeholders its actual post-resize vision grid produces.

Changelog

• Preserve fixed_512 as the compatibility default and enable processor mode only in the shipped Nemotron Omni VALOR Energon recipe.
• Reuse the public Nemotron Omni processor's aspect-preserving video pixels, imgs_sizes, and reported token counts instead of reimplementing its resize policy.
• Derive prompt-expansion counts independently for every tubelet while keeping logical num_image_tiles metadata separate.
• Validate frame/grid metadata and reject processor-driven sizing with the deprecated LLaVA collapse/expand contract.
• Cover 252- and 242-token rectangular grids, ragged packing, incomplete tubelets, text-only temporal batches, config propagation, and a real rectangular RADIO forward.
• Document the new mode and its compatibility boundary. Direct-HF behavior remains unchanged.

GitHub Actions CI

Focused validation on one H100:

• uv run --no-sync pre-commit run --all-files
• A real VALOR MP4 through the locked public processor: ten 384x672 frames, five 252-token tubelets, and 1,260 placeholders matching 1,260 expected vision rows.
• Focused Nemotron Omni builder, collator, Energon, recipe, utility, and rectangular RADIO GPU tests: 159 passed.
• git diff --check

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?
• Does the PR affect components that are optional to install? No.

Additional Information

This PR intentionally scopes dynamic temporal-video sizing to the canonical Energon path. The fixed mode remains available for compatibility, and the deprecated LLaVA contract remains fixed-resolution.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.