[nemotron] Add NemotronH (Mamba2-Transformer hybrid MoE) text-LLM sup…

[YunchaoYang]

…port

Phase 1 of NVIDIA Nemotron-3-Nano-Omni-30B-A3B: text-only LLM with a 52-layer 3-way hybrid decoder.

• Architecture: single-purpose blocks per layer (23 Mamba2 SSM, 23 MoE FFN, 6 GQA attention). Pattern is configurable via NemotronHConfig.hybrid_override_pattern.
• Mamba2 mixer: chunked scan via mamba_ssm triton kernels with a PyTorch fallback; incremental decoding via selective_state_update. Group-wise gated RMSNorm (8 groups of 512), gate-first ordering to match HF.
• MoE: 128 experts, top-6 sigmoid routing with e_score_correction_bias, plus one always-on shared expert. Squared-ReLU activation. Float32 output accumulation. Batched expert_hit transfer to avoid per-iteration CPU-GPU sync.
• GQA attention: 32 query heads, 2 KV heads, full head dim, no RoPE (Mamba handles position).
• HuggingFace interop: bidirectional state dict conversion handles language_model.backbone.* key prefix, fused projections, expert stacking, and GQA head layout.
• Tensor parallelism: shard_specs for attention, embedding, MoE (column- sharded inner FFN + row-sharded output + TP all-reduce), and final_proj.
• Asset card and composition/models.py registration.
• 60/60 unit tests pass (config, decoder layer, Mamba2 mixer, MoE).

What does this PR do? Please describe:
A summary of the change or the issue that is fixed.

Fixes #{issue number}

Does your PR introduce any breaking changes? If yes, please list them:
List of all backwards-incompatible changes.

Check list:

• Was the content of this PR discussed and approved via a GitHub issue? (no need for typos or documentation improvements)
• Did you read the contributor guideline?
• Did you make sure that your PR does only one thing instead of bundling different changes together?
• Did you make sure to update the documentation with your changes? (if necessary)
• Did you write any new necessary tests?
• Did you verify new and existing tests pass locally with your changes?
• Did you update the CHANGELOG? (no need for typos, documentation, or minor internal changes)