feat(bench): add local-coder benchmark harness

[phaedrus-on-git]

Summary

• adds a streamed devrail-router bench command for fixed OpenAI-compatible chat benchmark cases
• records request ID, HTTP status, first SSE event latency, duration, response bytes, streamed usage tokens, and first content sample as JSONL
• includes default local-coder cases and benchmarking docs for llm-srv

Validation

• go test ./...
• make check
• make package-smoke GOOS=linux GOARCH=amd64
• live sample against http://llm-srv-01.mfsoho.linkridge.net:18080/v1 with -model local-coder -max-tokens 256: 3/3 cases returned 200, first_event_ms 513-690, duration_ms 4024-4206