TGI per-request speculated-token histogram

TGI main, router/src/server.rs plus backends/v3/src/{backend,queue,lib}.rs, read 2026-10-04. On by default: the Prometheus recorder is built unconditionally with PrometheusBuilder and an HTTP listener on prometheus_port, default 9000, so nothing has to be opted into. Label keys are method (prefill|decode) on the batch families, err on tgi_request_failure, and reason on tgi_batch_concat - there are no model or engine labels, so every series is fleet-wide. tgi_request_input_length is matched by Matcher::Full with 100 linear buckets up to max_input_tokens; generated_tokens and max_new_tokens likewise up to max_total_tokens, so these three ladders depend on deployment sizing rather than being fixed.