Support OpenAI Reponses format API forward

[zsh2401]

https://platform.openai.com/docs/api-reference/responses

[diva0078]

This is a useful direction. From the gateway side, Responses API support tends to reveal compatibility gaps faster than chat-completions because the shape touches streaming events, tool calls, output item types, usage accounting, and error mapping at the same time.

If this gets implemented, I’d be especially curious about three edge cases:

1. streaming + tool calls in the same response
2. provider fallback when one upstream returns a rate-limit or schema error
3. mapping non-OpenAI backends into Responses-style output items without losing retry/debug info

I’m comparing Claude/OpenAI-compatible proxy patterns for Qwen/DeepSeek/Kimi-style backends, and these are the parts that seem to break first in real use.