fix(openai): prime Images streams before upstream first byte

[jiwangyihao]

Summary

• Prime OpenAI Images stream=true responses after local validation/routing and before waiting for the upstream response header.
• Send an initial SSE comment and pre-upstream keepalive comments using gateway.image_stream_keepalive_interval.
• Stop the primer before service-level response writes, and use SSE error events for post-primer upstream errors/fallbacks.

Motivation

Images stream requests can spend a long time waiting for the upstream response header before any downstream byte is written. Reverse proxies/CDNs with first-byte or idle timeouts may close the client connection before the existing stream keepalive loop starts.

This change moves the downstream SSE opening point earlier for Images stream=true requests without changing non-streaming behavior.

Behavior

• stream=false is unchanged.
• stream=true Images requests receive SSE headers and an initial : comment before upstream headers arrive.
• Once the SSE response has been opened, later upstream HTTP errors or non-SSE fallbacks are written as SSE event: error events instead of JSON responses.
• The primer is stopped before service forwarding writes to the same response writer.

Tests

• C:/Users/34404/Documents/GitHub/workbench/toolchains/go/bin/go.exe test -tags unit ./internal/handler ./internal/service -run 'TestOpenAIImagesStreamPrimer|TestHandleOpenAIImagesStreamPrimerForwardError|TestShouldWriteOpenAIImagesStreamPrimerForwardErrorIgnoresPrimerComments|TestOpenAIGatewayServiceForwardImages_(APIKeyStream|OAuthStreaming|APIKeyStreamJSONAfterPrimer|APIKeyStreamHTTPErrorAfterPrimer|OAuthStreamHTTPErrorAfterPrimer)' -count=1 -timeout=120s
• C:/Users/34404/Documents/GitHub/workbench/toolchains/go/bin/go.exe test -tags unit ./internal/handler ./internal/service -count=1 -timeout=180s
• git diff --check origin/main...HEAD

Refs #2031