fix(executor): Extract only tool_arguments for MCPToolBlock input reshaping

The entire merged input_data dict (containing server_url, credentials,
selected_tool, etc.) was being assigned to tool_arguments instead of
just the tool_arguments sub-dict. This would cause validation failures
or MCP server rejections.

autogpt_platform/backend/backend/executor/manager.py

@@ -232,7 +232,7 @@ async def execute_node(
         input_data = _input_data.model_dump()
     elif isinstance(node_block, MCPToolBlock):
         _mcp_data = MCPToolBlock.Input(**node.input_default)
-        _mcp_data.tool_arguments = input_data
+        _mcp_data.tool_arguments = input_data.get("tool_arguments", {})
         input_data = _mcp_data.model_dump()
     data.inputs = input_data