Dev - #1

Problem:
- _call_llm_with_tools was called every iteration (up to 800 times)
- Each call logged full routing info using WARNING level
- Caused massive log spam with identical messages
- Model switches were being logged even when no switch occurred

Solution:
1. Track last logged state (_last_logged_task_type, _last_logged_model)
2. Only log when task type or model actually changes
3. Count and report total model switches at completion
4. Move detailed routing config to workflow start (log once)
5. Change from WARNING to INFO level (less noise)
6. Remove duplicate logs from _select_model_for_task

Benefits:
- Clean logs: Only see changes, not repetition
- Performance: Less I/O overhead
- Visibility: Model switch count shows optimization effectiveness
- Debugging: Still see all important state transitions

Example output (before):
  🔍 TASK ROUTING CHECK: ... (800 times)
  ✅ Switched model: ... (800 times)

Example output (after):
  🔄 Task routing configuration: ... (once at start)
  🔄 Task routing: code_generation → qwen3-coder:30b (only on change)
  ✅ Model switch HKUDS#1: qwen3:32b → qwen3-coder:30b (only on actual switch)
  📊 Total model switches during implementation: 15 (at end)