[REQUEST] Expose llama.cpp-style top-level `timings` and cached/draft token stats in OAI responses

[dfein38347g]

(Drafted with AI assistance.)

When consuming tabbyAPI behind a metrics-aware proxy (e.g. llama-swap), per-request speed and cache stats don't surface, because the stats live only in the non-standard usage object while the standard places consumers look are empty.

Problem

1. Speeds are only in usage. UsageStats carries prompt_tokens_per_sec / completion_tokens_per_sec, but there is no top-level timings (llama.cpp-style) or metrics (vLLM-style) object. Consumers that read top-level timings/metrics for prompt/prediction speed show "unknown" even though token counts populate.
2. Cached prompt tokens are dropped. The backend (exllamav3) finish chunk already reports cached_tokens, but get_usage_stats discards it when building UsageStats, so the OpenAI-standard usage.prompt_tokens_details.cached_tokens is never emitted.
3. Spec-decoding draft stats are dropped. The finish chunk carries draft_accept / draft_reject, but they are not surfaced anywhere in the response.
4. Prompt count/speed mismatch. usage.prompt_tokens reports the total (including cached), while prompt_tokens_per_sec is computed over new tokens only — so the count and the speed disagree.

This is solvable and is working for us. We run a patched tabbyAPI image that exposes these fields, and the metrics populate correctly in our setup. We have a working patch set against v1.4.2 that we're happy to share for reference; we're not asking to own the implementation.

General direction for a fix

• Emit a top-level timings object on /chat/completions and /completions (streaming + non-streaming) with llama.cpp-compatible keys: prompt_n, prompt_per_second, predicted_n, predicted_per_second, and draft_n / draft_n_accepted when spec-decoding is active.
• Add prompt_tokens_details: {cached_tokens} to UsageStats, populated from the backend finish chunk's cached_tokens.
• Add draft_tokens / draft_tokens_accepted to UsageStats from draft_accept / draft_reject.
• Decide on prompt_tokens semantics: report new-only (total − cached) to match llama.cpp, or keep total and document it.