refactor(server): decompose worker request processing

[Graffioh]

Summary

• Reduce worker_loop() to dequeueing and delegate one request lifecycle to process_job().
• Extract FlowKV, whole-prompt PFlash, upstream forwarding, cache/snapshot lifecycle, generation setup, token I/O, and API-specific response serialization into focused helpers.
• Keep cache/decode state together where splitting it further would add more plumbing than clarity.

Example

Prompt compression changes now stay in apply_flowkv_compression() or apply_pflash_compression() instead of sharing a 1,500-line worker with cache restore, decoding, and response construction.

This is a behavior-preserving refactor; request and response formats are unchanged.

Testing

• git diff --check
• Unit test pass

[cubic-dev-ai]

No issues found across 2 files

Not reviewed (too large): server/src/server/http_server.cpp (~2,873 lines) - if these are generated or fixture files, add them to ignored paths to exclude them from future reviews.

_{Re-trigger cubic}