feat(antigravity): 增强网关功能和 thinking 块处理

主要改进：
- 优化 thinking blocks 过滤策略，支持 Auto 模式降级
- 将无效 thinking block 内容转为普通 text
- 保留单个空白 text block，不过滤
- 重构配额刷新机制，统一与 Claude 一致
- 支持 cachedContentTokenCount 映射到 cache_read_input_tokens
- Haiku 模型映射到 Sonnet
- 添加 /antigravity/v1/models 端点支持
- countTokens 端点直接返回空值

backend/internal/pkg/antigravity/response_transformer.go

@@ -232,10 +232,18 @@ func (p *NonStreamingProcessor) buildResponse(geminiResp *GeminiResponse, respon
 		stopReason = "max_tokens"
 	}
 
+	// 注意：Gemini 的 promptTokenCount 包含 cachedContentTokenCount，
+	// 但 Claude 的 input_tokens 不包含 cache_read_input_tokens，需要减去
 	usage := ClaudeUsage{}
 	if geminiResp.UsageMetadata != nil {
-		usage.InputTokens = geminiResp.UsageMetadata.PromptTokenCount
+		cached := geminiResp.UsageMetadata.CachedContentTokenCount
+		prompt := geminiResp.UsageMetadata.PromptTokenCount
+		if cached > prompt {
+			cached = prompt
+		}
+		usage.InputTokens = prompt - cached
 		usage.OutputTokens = geminiResp.UsageMetadata.CandidatesTokenCount
+		usage.CacheReadInputTokens = cached
 	}
 
 	// 生成响应 ID