feat(logs): capture request/response payloads and redesign log viewer

[shuaijinchao]

Backend:

• Extend request_logs schema with method, path, request_headers, request_body, response_headers, response_body (SQLite + Postgres migrations via ensure_request_log_column / ALTER TABLE IF NOT EXISTS)
• Capture ingress method/path/headers/body across universal, gemini and embeddings proxy entrypoints; aggregate final JSON for streaming responses and persist as response_body
• Emit logs on all early-exit error paths (decode failure, no route, auth failure, upstream error, cache miss fallbacks) with full context
• Cache-hit paths now include complete request/response bodies
• Embeddings proxy: log every success/error branch and parse usage.prompt_tokens into input_tokens
• Split API: query_logs list strips heavy fields (NULL bodies/headers); new get_log(id) endpoint fetches full payload on demand
• Settings: DEFAULT_RETENTION_DAYS 30 -> 7, batch size 64 -> 32, cleanup interval 60s -> 600s; new log_record_payloads toggle (default true) to disable payload persistence

Frontend:

• Compact 7-column log list (Time / Status / Model / Protocol / Latency / Token / Type), left-aligned, row click opens detail
• New LogDetailDialog with meta header and four copy-enabled payload blocks (request headers/body, response headers/body) using lazy get_log fetch and pretty-printed JSON
• Token display with IN/OUT labels and K/M formatting (<1000 raw, <1M 1-decimal K, >=1M 2-decimal M)
• Type badge: green SSE / sky JSON replaces boolean stream column
• Settings: split Log Configuration into its own half-width card next to Proxy Configuration; rename to "Retention Period" + "Record Payloads" with HelpCircle tooltips; tighten Config Backup layout

FIX #62