feat(server): serve per-request inference metrics under --enable-metr…

[FantasticCode2019]

…ics-report

• the scheduler measures the prefill span itself (admission to the token sampled off the request's last prefill chunk) and ships it as a duration on that token's DetokenizeMsg: it is the one boundary a client timing SSE frames cannot see, and a duration keeps the wire free of any cross-process clock assumption
• the remaining timings come off the raw ack stream, not the emitted GenEvents: a reasoning or tool-call parser holds text back, sometimes to the end of the stream, so an event-timed TTFT runs long, and acks are the one path the streaming and buffered generators share
• cached_prompt_tokens reports the real prefix-cache hit whether or not --enable-cache-report is set; that flag governs the billing fields in usage, and gating the hit here would leave prefill_tokens_per_second computed over tokens that were never forwarded
• decode_tokens_per_second divides by decode_tokens - 1, as vLLM's and sglang's serving benchmarks do: the first token falls out of prefill and spans no decode interval
• served on /v1/chat/completions, /v1/messages and /v1/responses; /v1/completions does not go through the shared generation core and is left out

Closes #503