feat: Implement multi-model LLM orchestration with LLMRouter

[FreecoDAO]

Summary

This PR implements a multi-model LLM orchestration layer that enables hybrid LLM usage for different tasks (e.g., using Claude for planning and Qwen for execution).

Changes

• New File: app/llm_router.py - Core LLMRouter class for dynamic model selection
• Modified: app/llm.py - Added ask_sync synchronous wrapper method
• Modified: app/agent/base.py - Updated to use llm_router for model initialization
• Modified: app/flow/planning.py - Updated to use llm_router for planning tasks
• Modified: README.md - Added multi-model configuration examples

Features

• Dynamic model selection based on task type or agent name
• Singleton pattern for efficient model management
• Backward compatible with existing single-model configurations
• Fallback to default model when specific model not configured

Configuration Example

[llm.planning]
model = "claude-3-5-sonnet"
base_url = "https://api.anthropic.com/v1"
api_key = "sk-ant-..."

[llm.executor]
model = "qwen-max"
base_url = "https://api.qwen.com/v1"
api_key = "sk-qwen-..."

Benefits

• Improved reasoning quality by using specialized models for different tasks
• Cost optimization by using cheaper models for simple tasks
• Flexibility to experiment with different model combinations

Testing

All changes have been tested for:

• Code consistency with existing patterns
• Security (no hardcoded secrets or dangerous functions)
• Backward compatibility