feat(mm-routing): support Kimi K3 image requests via llm-multimodal 1.9

[furionw]

Why

Kimi K3 multimodal KV routing needs the K3 vision spec added in llm-multimodal 1.9. Dynamo currently pins 1.7, so K3 falls back to text-prefix routing and cannot distinguish requests by image content.

Depends on #13059 for the vLLM 0.27.1 runtime support. Tracks DIS-2678.

What Change

• Bump llm-multimodal to 1.9 and llm-tokenizer to 1.6, including compatible Tokio, Mio, and Velo pins.
• Reproduce K3 worker routing tokens: media wrapper, original dimensions, and content-derived pad run.
• Add frontend-to-vLLM-KV-event hash parity coverage and advertise K3 in the support matrix.

Test Plan

• Two TP=8 workers; A/B share text and dimensions, differing only in image pixels.
• Each request has 6,184 prompt tokens; 1,536-token KV blocks yield four full blocks plus a 40-token tail.
• Cold A → worker 3236652464029061 at 0/5: none of five routing blocks overlapped.
• Cold B → worker 5439765509436224 at 0/5: independently seed B's image-specific cache.
• Warm A ×4 → A's worker at 4/5: four full blocks overlap, totaling 6,144 cached tokens.
• Warm B ×4 → B's worker at 4/5: same cache reuse; the partial tail is not reusable.

Dependency

• llm-multimodal 1.9 supplies Kimi-K3 support and requires llm-tokenizer 1.6 plus Tokio 1.52.3 or newer.
• Dynamo and Velo 0.1 pin Tokio 1.48, while Tokio 1.52 requires Mio 1.2 or newer.
• Tokio 1.52.3, Mio 1.2, and Velo 0.3 are therefore required resolver-compatible bumps, not unrelated upgrades.
• The remaining lockfile changes are transitive updates from the multimodal, tokenizer, and consolidated Velo dependency graphs.

[github-actions]

🌿 Fern Docs Preview: https://nvidia-preview-01a00c5c-7b48-746f-a8b6-4c36434b55ca.docs.buildwithfern.com/dynamo/dev

[furionw]

duplicate of #13371