feat: optimize tool selection middleware to cache and reuse tool selection per agent run

- Refactor MoviePilotToolSelectorMiddleware to perform tool selection once per agent execution and cache the result in state, avoiding redundant LLM calls for each model round.
- Add abefore_agent to select tools at the start of agent execution and store selected tool names in state.
- Update awrap_model_call to reuse cached tool selection from state for subsequent model calls.
- Enhance test coverage for tool selection caching and reuse logic.
- Improve error logging in skill version extraction.

app/agent/__init__.py

@@ -451,6 +451,7 @@ class MoviePilotAgent:
                         model=non_streaming_model,
                         max_tools=max_tools,
                         always_include=always_include_tools,
+                        selection_tools=tools,
                     )
                 )