feat(multimodal): graph Qwen2.5-VL custom encoder

[furionw]

Based on #11037.\n\n## Why\n\nDynamo’s custom encoder path can batch and cache image embeddings, but Qwen2.5-VL derives rotary and window metadata dynamically while segmented SDPA prevents efficient graph capture. Static grid metadata plus variable-length FlashAttention makes its complete 32-block vision tower capturable while preserving native vision parity. A separate performance experiment pairs the 3B vision tower with Qwen2.5-1.5B by truncating its completed 2048-wide output to 1536; this is untrained and makes no quality or model-parity claim. The Qwen3 encoder remains available for comparison.\n\n## Effect\n\nPerformance-only closed-loop H100 image sweep, 1,000 requests per cell:\n\n| Concurrency | TTFT avg (ms) | E2E avg (ms) | Throughput (req/s) |\n| ---: | ---: | ---: | ---: |\n| 8 | 102.8 | 501.2 | 15.865 |\n| 16 | 549.8 | 897.3 | 17.688 |\n| 32 | 1109.9 | 1519.6 | 20.734 |\n\n## What Change\n\n- Graph Qwen2.5 rotary, window attention, merger, and output reordering across static buckets.\n- Preserve Qwen3, caches, coalescing, pinned transfers, and bounded micro-batching.\n- Add audited H100 QPS and closed-loop concurrency benchmark workflows.\n\n## Test Plan\n\n- 36 focused H100 tests and all targeted pre-commit hooks passed.\n- Graph parity/allocator verification and audited 3,000-request concurrency sweep passed.

[github-actions]

🌿 Fern Docs Preview: https://nvidia-preview-019f5e77-6c97-768d-be3c-875f7e24ba5e.docs.buildwithfern.com/dynamo/dev

[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.