v2.8.0 — vllm-cuda audio-LLM microservice

Add vllm-cuda: Python FastAPI supervisor wrapping a single `vllm serve`
subprocess. Ships Qwen3-ASR-1.7B + Voxtral-Mini-3B with transcribe and
chat-with-audio endpoints (4 LiteLLM aliases). Only one model resident
at a time; LiteLLM resource_manager cuda-vllm group handles eviction.

Wrapper rewrites multipart response_format=verbose_json -> json before
forwarding to vllm (LiteLLM forces verbose_json; vllm doesn't accept
it). 9/9 live e2e tests pass.

Granite-Speech-3.3-2B was tried and dropped — its rel_pos_emb tensor is
architecturally ~9.5 GiB, incompatible with 12GB cards in a shared-GPU
deployment.