fix(proxy): register images/variations as inference routes for budget enforcement

LiteLLMRoutes.openai_routes drives RouteChecks.is_llm_api_route(); the
variation paths were absent, so a key explicitly granted either path via
allowed_routes bypassed the virtual-key max-budget check and the common
team/user budget gates - a provider-backed request could keep running
after the budgets were exhausted. Registered both paths so the standard
gates run, matching edits and generations.

litellm/proxy/_types.py

@@ -253,6 +253,9 @@ class LiteLLMRoutes(enum.Enum):
         # image edit
         "/images/edits",
         "/v1/images/edits",
+        # image variation
+        "/images/variations",
+        "/v1/images/variations",
         # video generation
         "/videos",
         "/v1/videos",