FEAT: report cache hit tokens across vLLM, SGLang and MLX

[qinxuye]

vLLM and SGLang currently discard the engine's per-request cache hit count when building completion responses, so clients cannot see native prefix-cache or Xavier cache reuse. This PR exposes that count as usage.prompt_tokens_details.cached_tokens in completions and chat completions, including streaming responses. The existing Responses API adapter forwards it as usage.input_tokens_details.cached_tokens.

MLX already reports cached tokens, but its continuous batching path does not emit the final usage-only chunk requested by stream_options.include_usage. This PR adds that chunk and verifies actual Xavier reuse through the MLX model adapter.

In heterogeneous P/D, the router forwards the decode engine's usage unchanged. The count includes KV imported from the prefill engine, including KV computed for this request; it does not separately measure historical prefix-cache hits on the prefill engine.

• Preserve the count in the final finish chunk and the optional choices=[] usage chunk across all three engines.
• Pass through existing engine counts without additional cache queries, RPCs, or cache scheduling changes. SGLang's aggregate count is used directly, avoiding double counting its device/host/storage breakdown; vLLM counts the cached prompt once when generating multiple outputs.
• Omit prompt_tokens_details when older engines return no count, while preserving reported zeroes. The Responses API retains its existing zero fallback.
• Document the fields and streaming option in the client guide, with translations for all nine maintained locales.

Validation:

• 33 cached-token regression cases passed on macOS and Linux with vLLM 0.21.0: cache hit/miss, missing/None counters, completions, chat conversion, Responses conversion, streaming finish/usage chunks, and multiple vLLM outputs.
• macOS regression slice: 465 passed, 9 skipped, 10 deselected. This covers the engine adapters, shared chat utilities, MLX/Xavier, and Responses API. The deselections are three full-model launch tests and seven failures independently reproduced on the base revision.
• Real Metal MLX/Xavier tests passed in both streaming and non-streaming modes: cold requests report 0, fresh remote-prefix reuse and P/D handoff report 69 cached tokens out of 70 prompt tokens, with identical generated text.
• Linux runtime slice with vLLM 0.21.0: 341 passed, 8 skipped, plus three failures reproduced on the base revision (listed below).
• Heterogeneous P/D regression slice: 188 passed, including eight new router cases covering chat/completions, streaming/non-streaming and GPU/host modes. The router preserves D usage without adding or replacing it with P usage. Four existing real Metal host-import cases also passed.
• Review follow-up: 211 passed on macOS and 33 passed on Linux with vLLM 0.21.0. Missing/None counters are tested only for vLLM/SGLang, including the Responses zero default. Six tools-streaming hit/miss cases cover the Qwen parser, usage-only chunk and Responses mapping across all three engines. The P/D count semantics are documented and rendered in English and all nine locales.
• Full pre-commit run --all-files and commit hooks passed.
• All nine modified catalogs passed msgfmt --check --check-format; python doc/build_i18n.py --all passed. The changed page built in English and all nine locales, and both translated paragraphs were verified in the rendered HTML.

Existing failures reproduced on base cb6c1ad2694515f42b23abd61ddbf51758226389:

• Two vLLM tool-stream tests have stale expected chunks (test_async_to_tool_completion_chunks_without_thinking and test_async_to_tool_completion_chunks_with_parser).
• Five vLLM configuration tests fail on macOS without vLLM installed because they compare VLLM_VERSION=None; they pass in the installed vLLM 0.21.0 Linux environment.
• The old Linux distributed-executor integration test imports vllm.executor.executor_base, removed by vLLM 0.21.0.

No full-model serving benchmark, full heterogeneous serving run, or real SGLang/Xavier GPU run was performed for this response-mapping change; the available SGLang environment is 0.5.20, below Xavier's 0.5.21 requirement.

[rogercloud]

Minor

• doc/source/user_guide/client_api.rst:99 — P/D semantics of cached_tokens are undocumented (decode reports imported KV as cached, so it is close to prompt_tokens even for cold requests).
• xinference/model/llm/tests/test_cached_token_usage.py:212 — MLX MISSING/None cases test a state production MLX cannot produce.
• xinference/model/llm/tests/test_cached_token_usage.py:218 — no coverage for the tools path or the Responses zero default.

Blocking: no — recommended event: APPROVE