refactor(ws): reuse marshaled payload JSON for write and prewarm

[penwyp]

Summary

forwardOpenAIWSV2 can end up marshaling the same request payload twice: once to resolve payload_bytes for WS schema logging and again when the payload is written to the websocket. performOpenAIWSGeneratePrewarm had the same pattern: it built JSON bytes for logging but still passed the map to WriteJSONWithContextTimeout, forcing another marshal on write.

This PR narrows the WS optimization work to one focused change: marshal the payload once, reuse the same JSON bytes for the actual WS write/prewarm write, and fail closed if marshaling fails.

Changes

• backend/internal/service/openai_ws_forwarder.go
  • cache the marshaled request payload in forwardOpenAIWSV2 and reuse it for websocket writes via json.RawMessage
  • marshal the prewarm payload once in performOpenAIWSGeneratePrewarm, log the byte size, and write the same JSON upstream
  • return explicit fallback errors when write/prewarm marshaling fails instead of attempting to write a partial request
• backend/internal/service/openai_ws_forwarder_success_test.go
  • add regression coverage that the forwarded WS payload JSON is equivalent to direct json.Marshal(payload)
  • add fail-closed coverage for write-path marshal failures
  • add regression coverage that the prewarm payload JSON is equivalent to direct json.Marshal(prewarmPayload)
  • add fail-closed coverage for prewarm marshal failures

Test plan

• cd backend && go test ./internal/service/...
• cd backend && go test ./...