feat: add image generation for OpenAI DALL-E and Gemini Imagen

New `/v1/images/generations` endpoint proxies DALL-E 2/3 (OpenAI)
and Imagen 3 (Gemini). Same auth/logging as chat completions.

- Add ImageGenerationRequest/Response models
- Extend Provider interface with ImageGeneration()
- OpenAI: forward to /v1/images/generations
- Gemini: call /v1beta/models/{model}:predict, map OpenAI params
- Circuit breaker wraps image gen like chat completions
- Model routing: dall-e* -> openai, imagen*/gemini* -> gemini
- Unsupported providers (deepseek/moonshot/grok/ollama) return error
- Fix pre-existing CachedContentTokenCount bug in StreamGemini

internal/providers/helpers.go

@@ -256,9 +256,10 @@ func StreamGemini(ctx io.ReadCloser, ch chan<- *models.ChatCompletionStreamRespo
 					FinishReason string `json:"finishReason"`
 				} `json:"candidates"`
 				UsageMetadata struct {
-					PromptTokenCount     uint32 `json:"promptTokenCount"`
-					CandidatesTokenCount uint32 `json:"candidatesTokenCount"`
-					TotalTokenCount      uint32 `json:"totalTokenCount"`
+					PromptTokenCount        uint32 `json:"promptTokenCount"`
+					CandidatesTokenCount    uint32 `json:"candidatesTokenCount"`
+					TotalTokenCount         uint32 `json:"totalTokenCount"`
+					CachedContentTokenCount uint32 `json:"cachedContentTokenCount"`
 				} `json:"usageMetadata"`
 			}