fix(gateway): normalize DeepSeek reasoning_effort 'max' to 'xhigh'

pi maps defaultThinkingLevel 'xhigh' → 'max' for deepseek-v4-pro via
thinkingLevelMap, but normalizeOpenAIReasoningEffort did not recognize
'max', causing all reasoning_effort to be dropped (0% fill rate since
traffic switched from claude to deepseek on May 1).

Add 'max' → 'xhigh' mapping to match Claude's NormalizeClaudeOutputEffort
behavior.

backend/internal/service/openai_gateway_service.go

@@ -7311,7 +7311,7 @@ func normalizeOpenAIReasoningEffort(raw string) string {
 		return ""
 	case "low", "medium", "high":
 		return value
-	case "xhigh", "extrahigh":
+	case "xhigh", "extrahigh", "max":
 		return "xhigh"
 	default:
 		// Only store known effort levels for now to keep UI consistent.