feat(recipe): refine Qwen3.5 and Qwen3.6 VL pretraining

[cuichenx]

What does this PR do ?

Refines the architecture-shared Qwen3.5/Qwen3.6-VL 35B-A3B pretraining recipes and records the verified Qwen3.6 full-model DataComp result.

Changelog

• Expose the maintained full-pretraining recipe only as qwen35_vl_35b_a3b_pretrain_config and use that canonical name in the verification card.
• Use Megatron-Core's canonical gdn selector for the shared Qwen3.5/Qwen3.6 bridge and provider configuration.
• Update the Qwen3.6 verification card's functional pretrain evidence to 135.610 model TFLOP/s/GPU (approximately 136) on 16 H100 GPUs with natural routing and pinned public DataComp provenance.
• Keep checkpoint resume explicitly unverified and omit the standalone performance section pending separate evidence.
• Expand the MoE VLM performance-tuning skill with fit-first TP/EP, recompute, HybridEP, memory, and MBS guidance.
• Make the no-performance-result card disclaimer describe evidence recorded by a card, while retaining validator compatibility for existing cards.

GitHub Actions CI

Focused local validation in nvcr.io/nvidia/nemo:26.06:

• Qwen3.6 model-card validator: passed.
• Model-card validator and Qwen VL recipe unit tests: 160 passed.
• uv run --no-project pre-commit run --all-files: passed.
• git diff --check: passed.
• Fresh independent final-diff review: no findings after all actionable feedback was addressed.

Before your PR is "Ready for review"

Pre checks:

• Read and followed the contributor guidelines.
• Added focused tests for the canonical recipe export and validator behavior.
• Updated the model verification card and relevant skill documentation.
• Confirmed this PR does not change optional-library import behavior.

Additional Information

• No related issue.
• The card intentionally classifies the measured DataComp run as functional pretrain evidence rather than a standalone performance item.

[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.