fix: handle dict type for server_tool_use in Usage model

Fix Pydantic serialization warning when API responses contain
server_tool_use as a dict instead of ServerToolUse object.

Some API providers (e.g., NEXUS, custom Anthropic-compatible endpoints)
return server_tool_use as a plain dict with fields like:
{
  "web_search_requests": 0,
  "tool_search_requests": null
}

This caused Pydantic serialization warnings:
  PydanticSerializationUnexpectedValue(Expected `ServerToolUse` -
  serialized value may not be as expected)

Changes:
- Modified Usage.__init__ to accept Union[ServerToolUse, dict] for server_tool_use
- Added conversion logic to transform dict to ServerToolUse object
- Follows same pattern as prompt_tokens_details and completion_tokens_details

This maintains backward compatibility while supporting API responses
that return dict format for server_tool_use field.

litellm/types/utils.py

@@ -1428,7 +1428,7 @@ class Usage(SafeAttributeModel, CompletionUsage):
         completion_tokens_details: Optional[
             Union[CompletionTokensDetailsWrapper, dict]
         ] = None,
-        server_tool_use: Optional[ServerToolUse] = None,
+        server_tool_use: Optional[Union[ServerToolUse, dict]] = None,
         cost: Optional[float] = None,
         **params,
     ):
@@ -1521,6 +1521,14 @@ class Usage(SafeAttributeModel, CompletionUsage):
                     "cache_creation_input_tokens"
                 ]
 
+        # Handle server_tool_use - convert dict to ServerToolUse object if needed
+        _server_tool_use: Optional[ServerToolUse] = None
+        if server_tool_use is not None:
+            if isinstance(server_tool_use, dict):
+                _server_tool_use = ServerToolUse(**server_tool_use)
+            elif isinstance(server_tool_use, ServerToolUse):
+                _server_tool_use = server_tool_use
+
         super().__init__(
             prompt_tokens=prompt_tokens or 0,
             completion_tokens=completion_tokens or 0,
@@ -1529,8 +1537,8 @@ class Usage(SafeAttributeModel, CompletionUsage):
             prompt_tokens_details=_prompt_tokens_details or None,
         )
 
-        if server_tool_use is not None:
-            self.server_tool_use = server_tool_use
+        if _server_tool_use is not None:
+            self.server_tool_use = _server_tool_use
         else:  # maintain openai compatibility in usage object if possible
             del self.server_tool_use