feat(server): add per-request table log for chat completions

[cdryzun]

Summary

Adds a table-formatted request log for /v1/chat/completions. Each call prints one row to stdout with sequence number, time, model, mode (stream/sync), status, time-to-first-token, output token count, tokens/sec and total duration. This makes slow or throttled requests visible without digging through raw nginx access logs.

Why

The service currently logs only its startup line and any upstream transport errors. There is no per-request visibility into latency or token usage, which makes it hard to tell when a specific model or account is underperforming. The table gives an operator a quick per-request picture.

What changed

• internal/server/logging.go (new): a chatStatsReader that wraps the upstream SSE stream. It counts streamed output tokens (by UTF-8 runes) and records the time to the first content frame. A logChatRow helper prints the aligned table row.
• internal/server/handler.go: wires the logger into the stream and non-stream paths. Error paths (bad request, upstream parse failure, all accounts unavailable) now also emit a row with the real HTTP status instead of being silent. /healthz is excluded, and chat completions no longer duplicate the plain one-line log.
• internal/server/logging_test.go (new): unit tests for SSE token counting (including reasoning_content), usage-field precedence, TTFB baseline, and model parsing.
• Dockerfile: sets GOPROXY to goproxy.cn so the container build can fetch modules behind the GFW. Unrelated to logging, included separately for completeness.

Notes

• Streamed token counts are estimates based on character count; non-streamed calls use the exact completion_tokens from the upstream usage field. The two are not directly comparable.
• Time-to-first-token is measured from when the request enters the handler, not from when the upstream connection is established.