fix: preserve the stream property for chat/completions calls

[SasSwart]

In the openai-go SDK, a conflict between a recent optimization and the way in which options are set resulted in an issue whereby request bodies for chat/completions calls were being discarded.

This PR fixes the issue by taking full control of the request body handling. We now use the WithRequestBody and WithJSONSet options ourselves to ensure that the request payload is sent exactly as we need it.

Integration tests have been added to guard against future regression.

Related to internal Slack thread.

[ssncferreira]

LGTM
Can you add a simple description to the PR and link it to the slack thread for future reference?

[ssncferreira]

We can actually use responses.ResponseNewParams:

Suggested change

	var reqBody map[string]any
	if err := json.Unmarshal(body, &reqBody); err != nil {
	return fmt.Errorf("request should be valid JSON: %w", err)
	}
	// Verify required fields for OpenAI responses
	// Note: Using map here since there's no specific SDK type for responses endpoint
	model, ok := reqBody["model"]
	if !ok || model == "" {
	return fmt.Errorf("model field is required but missing or empty")
	}
	return nil
	var req responses.ResponseNewParams
	if err := json.Unmarshal(body, &req); err != nil {
	return fmt.Errorf("request should unmarshal into ResponseNewParams: %w", err)
	}
	// Collect all validation errors
	var errs []string
	if req.Model == "" {
	errs = append(errs, "model field is required but empty")
	}
	if len(errs) > 0 {
	return fmt.Errorf("validation failed: %s", strings.Join(errs, "; "))
	}
	return nil

[ssncferreira]

I can add this as a follow-up 👍