[TRTLLM-14880][feat] qualify Qwen3 dense for MX

[chienchunhung]

@coderabbitai summary

Description

Add the second dense-model qualification profile for ModelExpress post-transform
weight reception. This PR reuses the exact runtime-constraint machinery from
#16974 and keeps the production change to a Qwen3-specific transform ABI plus an
exact Qwen3ForCausalLM registry row.

The initial profile covers a Qwen3-8B-like dense configuration and explicitly
qualifies the Q/K-normalized attention lifecycle, fused QKV and gate-up layouts,
default fused QK-norm/RoPE path, and untied language-model head.

Dependency

This PR is stacked on #16974 and must merge after it. Because GitHub cannot use a
fork-only branch as the upstream PR base, the draft targets main and will
temporarily display the parent changes until #16974 merges.

#16458  MX post-transform qualification foundation [merged]
`-- #16974  Qwen2 dense model-family support [draft, CI running]
    `-- this PR  Qwen3 dense model-family support [draft]

#17029 is a parallel ModelExpress transport-delegation refactor. It does not add
model-family profiles, but it overlaps the loader integration and may require a
downstream sync if it lands before this stack.

Qualified Scope

• Exact root/config identity: Qwen3ForCausalLM / qwen3
• Target-model post-transform transfer only
• Dense BF16 with unquantized weights and KV cache
• TRTLLM attention with default fused QK-norm/RoPE
• TP1 and TP2; PP1, CP1, and EP1
• Untied embeddings and no speculative mode or separately loaded draft model

Unsupported runtime variants fail closed through the shared dense-Qwen support
matrix, including FP16, quantization, alternate attention backends, TP greater
than 2, PP/CP/EP expansion, LoRA, sparse attention, attention DP, multi-node
execution, tied embeddings, YaRN, unfused RoPE, and speculative decoding.

Test Coverage

• Added full-load versus staged-load lifecycle equivalence for Qwen3 dense TP1.
• Added TP2 rank-local equivalence for ranks 0 and 1.
• Added Q/K norm, fused-layout, transform-guard, parameter/buffer, and
  deterministic output probes, with an unregistered-root negative control.
• Extended every shared dense-Qwen runtime rejection case to Qwen3.
• Extended SourceIdentity transform-ABI binding coverage.
• All changed-file pre-commit hooks pass.
• Python 3.12 compilation passes for all changed Python files.

The focused PyTorch tests and real Qwen3 ModelExpress donor/receiver GPU E2E have
not run locally because this checkout has no PyTorch/Transformers runtime. They
remain required before the draft is marked ready.

PR Checklist

Please review the following before submitting your PR:

• PR description clearly explains what and why. If using CodeRabbit's summary, please make sure it makes sense.

• PR Follows TRT-LLM CODING GUIDELINES to the best of your knowledge.

• Test cases are provided for new code paths (see test instructions)

• If PR introduces API changes, an appropriate PR label is added - either api-compatible or api-breaking. For api-breaking, include BREAKING in the PR title.

• Any new dependencies have been scanned for license and vulnerabilities

• CODEOWNERS updated if ownership changes

• Documentation updated as needed

• Update tava architecture diagram if there is a significant design change in PR.

• The reviewers assigned automatically/manually are appropriate for the PR.

• Please check this after reviewing the above items as appropriate for this PR.

GitHub Bot Help

To see a list of available CI bot commands, please comment /bot help.