Add per-request prompt cache files to server

[Quiet-Node-io]

Summary

Adds optional per-request disk prompt-cache support to the OpenAI-compatible server by accepting prompt_cache_file / prompt-cache-file in chat and text completion requests.

The server already supports an in-memory prompt cache and the CLI path already supports --prompt-cache-file; this change lets server callers use the same disk persistence primitive without switching away from /v1/chat/completions.

Behavior

• prompt_cache_file loads a matching disk prompt cache into an isolated request-local LRUPromptCache before prefill.
• The server saves a prompt cache file after the first generated token so future matching-prefix requests can reuse the prefix.
• Saved disk caches are trimmed by one prompt token to avoid exact-cache-hit requests entering generation with an empty remaining prompt.
• Cache-file requests are routed through the single-request path, not batching, until batch semantics are defined for per-request disk files.
• disable_prompt_cache forces an isolated request-local cache and suppresses disk load/save even if a caller sends a cache file path.
• Requests without these fields keep the existing non-cache server path.

Validation

• python -m py_compile mlx_lm/server.py