feat(vllm): complete exact MM routing for generate

[furionw]

Based on #11588.

Why

The frontend can project exact image identity into KV-routing tokens, but workers must advertise compatible support and normalize vLLM KV events to the same identity. The backend also needs to execute tokenized /inference/v1/generate requests without changing existing chat/completions or P/D behavior.

What Change

• Add native Generate handling to unified and legacy vLLM paths.
• Advertise exact-MM routing only when the worker resolves an image placeholder token.
• Normalize multimodal KV-event identifiers and preserve LoRA/cache namespaces.
• Propagate prompt embedding metadata through backend bindings.

Test Plan

• Run pre-commit, release Rust builds, 24 focused Python tests, 7 KV-event parser tests, and the real vLLM engine integration test.
• Run salted and unsalted two-worker Qwen3.5 multimodal routing E2E: all eight requests returned 200; distinct images warmed separate workers and repeats matched 2/3 cached blocks on the same worker.

[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.

[datadog-official]

🎯 Code Coverage (details)
• Patch Coverage: 100.00%
• Overall Coverage: 8.10% (-1.70%)

_{This comment will be updated automatically if new data arrives.
🔗 Commit SHA: ce7a9b2 | Docs | Give us feedback!}

[biswapanda]

Repair stack pushed at ce7a9b2fdf (still draft and intentionally based on #11588 until #11588 lands).

Blocker repairs included:

• preserved current-main native vLLM generate behavior and deleted unified-backend state
• completed exact multimodal routing event/hash plumbing
• preserved routed-expert, prompt-logprob, KV-transfer, usage, LoRA admission, and cancellation parity
• rejected invalid top_k < -1 at the frontend boundary
• made routed-expert serialization failures fail visibly instead of silently dropping metadata
• added the required core test marker

Local validation: Generate Rust tests 73 passed; KV-router wire tests 32 passed; Ruff, compileall, cargo fmt, and diff checks passed. Python vLLM collection is locally blocked only because the installed native extension predates FrontendExtensionContext; CI will exercise the built extension.

[biswapanda]

/ok to test ce7a9b2