feat(model): add canonical Nemotron Omni model

[aroshanghias-nvd]

What does this PR do?

Add a dedicated NemotronOmniModel and conversion contract for processor-expanded multimodal sequences without using the historical Nemotron Omni LLaVAModel collapse/expand path.

Fresh review requested: The branch history has been replaced with a main-based implementation of the reviewed data-boundary packing design. The earlier model-owned-packing implementation is obsolete; please review this series from the beginning.

The ownership boundary is now:

• The data pipeline packs processor-expanded token sequences and constructs PackedSeqParams. MBridge's canonical collator provides this behavior for native MBridge training; external callers such as NeMo-RL may provide the same caller-packed THD contract from their own data pipeline.
• NemotronOmniModel receives the full packed THD sequence, builds and inserts media embeddings, and only then selects the context-parallel portion of the merged embedding sequence.
• Packing does not run inside model forward.

Changelog

• Add NemotronOmniModel, reusing the existing RADIO encoder, multimodal projector, and Nemotron-H language model components.
• Preserve expanded image placeholders through the canonical collator, with packing performed at the data boundary.
• Keep the legacy LLaVA collator and provider/bridge compatibility path explicit for compatible legacy checkpoints.
• Add the caller-packed THD model contract needed to insert media embeddings before context-parallel slicing.
• Register canonical Nano and Super Omni HF architectures with AutoBridge, with versioned contracts that reject ambiguous legacy checkpoints.
• Add canonical import/export mappings for the top-level model namespace and reverse MTP QKV mapping support for Nemotron-H.
• Add conversion, model-forward, media-alignment, packing, Mamba-boundary, activation, and compatibility tests.

Validation

• Ruff lint and formatting checks passed.
• Focused unit and functional validation passed, including the two-rank caller-packed context-parallel path.
• The corresponding NeMo-RL distributed suite passed CP2 checkpoint, TP/SP, TP2+CP2, padded multi-row attention, and PP2 coverage against this exact MBridge head.
• Nano image-GRPO validation covering generation, training, repeated vLLM refit, checkpointing, and probability agreement is recorded in W&B.

Scope and follow-up

This PR implements the canonical image/text foundation used by the current Nano migration. Audio and video support is layered separately in #5007. Super/MTP integration and validation remain follow-up work. This implementation does not require changes to Megatron-LM source.

GitHub Actions CI

See the CI section in the contributing guide for how to trigger CI. An NVIDIA developer may need to approve and trigger CI for external contributors.

Before this PR is ready for review

• Read and followed the contributor guidelines.
• Added the necessary tests.
• Added or updated the necessary documentation.
• No new optional dependency is introduced.

[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.

[aroshanghias-nvd]

/ok to test 7a9e305

[cuichenx]

audio/video capabilities in #5007

[cuichenx]

/ok to test 124b19b

[cuichenx]

/ok to test f461eae