Add Qwen 3.8 27B Megatron support - #820

Summary

Adds qualified Megatron support for Qwen/Qwen3.8-27B by reusing the existing
Qwen 3.5/3.6 dense hybrid handler. Qwen 3.8 has the same model geometry and
weight schema, so this does not add a new handler or fork the linear-attention
implementation.

MTP remains deliberately disabled, consistent with the other Qwen 3.5-family
models in ART.

Differences from #810

This carries forward #810's small registry, TrainerRank, and focused-test
changes on current main, while resolving the minor qualification gaps found
when running the complete model-support workflow:

• bases the change directly on current main instead of the older
  codex/require-peft-0191 branch;
• pins the exact qualified Qwen 3.8 Hugging Face revision;
• selects the deterministic FP32 GDN reference by handler identity rather than
  literal qwen3.5 model-name text, which also fixes the registered Qwen 3.6
  aliases;
• records Qwen 3.8's exact measured B300 throughput fingerprint without
  replacing Qwen 3.5's calibration or performance floors;
• constrains the root and Megatron runtime to openai<3: OpenAI 3.x uses
  httpx2 internally and rejects ART's httpx 0.x Timeout/Limits objects
  before a request reaches the socket.

Validation

The complete default handler workflow passed from clean commit
cd61066789ed331614ce9ef6b1ef136dbb345251 on 8 B300s in 12m00s. All ten
mandatory stages ran through the normal overlapped scheduler. Optional
sensitivity variants were not requested (sensitivity_skipped=true).

• HF parity: 55/55 signals passed; output MAPE 0.004975%; worst gradient MAPE
  0.198038% (layers.2.linear_attn.dt_bias).
• Train/inference: passed first attempt; LoRA MAPE 7.80685% against the existing
  8.05% family limit; top-20 candidate-to-target KL 0.002315 against 0.003.
• Chat/tool template: 6/6 scenarios passed.
• Correctness: CP1 reference and TP2/CP2/PP2/VPP2 composition passed.
• Reduced 8-layer throughput stage: 69,435 isolated train tok/s, 64,852 E2E
  train tok/s, 4,053 accepted train tok/s, 1.0017 matched-core/isolated ratio,
  5.14% trainer underfeed, and no acceptance failures.
• Focused registry, TrainerRank, and workflow checks: 108 passed, 8 skipped;
  the final calibration resolver check was rerun after refreshing the contract.
• Ruff, formatting, type checking, and both lock consistency checks passed.

This supersedes the implementation proposed in #810.