fix(vllm): use engine-reported logical KV block sizes

[JulienDarve]

Summary

Use vLLM's initialized logical block size throughout the legacy Python backend. With physical block size 16 and DCP=2, full KV events contain 32 tokens; configuring Dynamo with 16 causes those events to be rejected.

• Fetch typed metadata through AsyncLLM.get_kv_cache_group_metadata() and cache the main-attention logical size for the publisher, base-model registration, and LoRA registration.
• Reject unavailable, invalid, or conflicting main-attention sizes. Preserve the no-cache case without deriving DCP or PCP multipliers.
• Add compact metadata coverage and one four-H100 DCP event-indexing regression using Qwen2.5-3B.

Validation

• 233 Python tests passed, including existing worker-factory and LoRA registration coverage.
• Existing TinyLlama KV-router end-to-end test passed on one RTX 6000 Ada: 84.77 seconds.
• Required pre-commit hooks passed.
• TP=4/DCP=2 configuration validated through vLLM. Multi-GPU execution, negative control, VRAM profiling, and nightly-test promotion for PR CI remain pending.

Tests used Python source from JulienDarve/vllm#3 at 20905fbeda0eb9760e1782d6ea5a0f96d4fe2457, with existing precompiled extensions.

Dependencies

Draft until the metadata API is available across supported CUDA, CPU, and XPU runtimes. The current vLLM 0.28.0 CUDA/CPU and 0.27.1 XPU pins do not provide this contract.

Related Issues

• DYN-3997
• Depends on DYN-4277 and [Feature][Frontend] Expose effective attention block size for DCP JulienDarve/vllm#3
• Supersedes fix(vllm): account for DCP in KV event block size #13653.

[JulienDarve]

Being implemented in the version bump PR: #15182