[docs] docs: Add model variant pages

[yaoyu-33]

Summary

Audit supported model variants after #3908 and add missing model-variant leaf pages where the repo clearly has bridge/example/recipe support but docs only had family/index coverage.

Added leaf pages for:

• Bailing: Ling 2.0
• Falcon: Falcon H1
• Gemma: Gemma
• GLM: GLM-4.7 / 4.7-Flash, GLM-5 / 5.1
• Kimi: Kimi K2, Kimi-K2.5-VL
• Llama: Llama 2
• MiniMax: MiniMax-M2 / M2.5 / M2.7
• Xiaomi-MiMo: Xiaomi-MiMo
• Qwen: Qwen3-MoE, Qwen3-Next, Qwen2-Audio, Qwen2.5-Omni, Qwen3-ASR
• Sarvam: Sarvam
• StepFun: Step-3.5-Flash, plus a new family index

Also updated family index toctrees, docs/models/README.md, docs/index.md, and the root README supported-model table for StepFun.

Audit notes

Sources checked:

• README.md supported models table
• docs/models/*/index.md and existing docs/models/*/*.md
• examples/models/* and examples/megatron_mimo
• src/megatron/bridge/models/*
• src/megatron/bridge/recipes/*

Intentionally kept grouped rather than split further:

• DeepSeek V2/V3/V4, Gemma 2/3/3-VL/4-VL, GLM-4.5/4.5V, GPT-OSS, Mistral/Ministral 3, Moonlight, OLMoE, Qwen3-VL, Qwen3.5/3.6-VL, Qwen3-Omni, and existing Nemotron pages already had leaf coverage.
• Llama 3 / 3.1 / 3.2 / 3.3 remain under llama3.md because they share the existing Llama 3 page and recipe family.
• Qwen2 / Qwen2.5 / Qwen3 dense variants remain under qwen.md because the existing page already covers the shared dense Qwen bridge and recipes.
• Nemotron H and Nemotron Nano v2 remain combined in nemotronh.md because the page explicitly covers both shared hybrid SSM-attention families.
• MiniMax-M2.5 / M2.7 remain grouped with MiniMax-M2 because the bridge source documents them as the same MiniMaxM2ForCausalLM architecture.
• GLM-4.7 / 4.7-Flash and GLM-5 / 5.1 are grouped in pairs because each pair shares the same example area and closely related bridge support.

Validation

• git diff --check passed.
• Focused changed-file local markdown link check passed: all changed-file local links resolve.
• Focused model toctree check passed: all docs/models/*/index.md toctree entries resolve.
• Requested command cd docs && uv run --no-sync sphinx-build --fail-on-warning --builder html . _build/html was blocked because the unsynced venv did not have sphinx-build installed.
• cd docs && uv run --group docs sphinx-build --fail-on-warning --builder html . _build/html was blocked by the local resolver/platform issue: nvidia-resiliency-ext==0.6.0 has no wheel for this host's manylinux_2_31_x86_64; available wheels are manylinux_2_39_*.
• Docs-tool-only build passed: cd docs && uv run --no-project --with sphinx --with myst-parser --with sphinx-autodoc2 --with nvidia-sphinx-theme --with sphinx-copybutton --with sphinxcontrib-mermaid sphinx-build --fail-on-warning --builder html . _build/html.
• Required exact pre-commit command uv run pre-commit run --all-files was blocked by the same nvidia-resiliency-ext platform resolver issue.
• Pre-commit-tool-only fallback passed: uv run --no-project --with pre-commit pre-commit run --all-files.

Unit tests were not run, per task instructions.

[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.

[claude]

Light Code Review

LGTM — docs-only PR adding model-variant leaf pages. All referenced implementation paths (bridge modules, recipes, example directories) verified to exist on this branch. No bugs, typos, or missing references found.
Suggested test cases: No perf tests impacted.