feat: improve codex chat completions format conversion and response handling

This commit enhances the Codex API implementation with several key improvements:

## Response Format Standardization
- Made `usage` field optional in `OpenAIChatCompletionResponse` to match OpenAI API spec
- Improved type safety by returning `OpenAIUsage` objects instead of raw dictionaries
- Added better null checking for response data to prevent crashes on malformed responses

## Reasoning Content Processing
- Simplified reasoning block handling to follow official OpenAI Response API format
- Removed complex signature tracking and custom XML wrapping
- Fixed reasoning content to stream properly without custom modifications
- Improved chunk processing with cleaner state management

## Streaming Response Improvements
- Enhanced streaming response detection with better header checking
- Fixed usage data serialization in streaming chunks using proper model dumping
- Improved error handling for edge cases in streaming responses

## Documentation Updates
- Updated README with comprehensive Codex API documentation
- Added clear limitations section for chat completions endpoint
- Included working model examples (gpt-5) and configuration samples
- Improved formatting consistency throughout documentation

## Model and Proxy Service Updates
- Enhanced model mapping validation and error handling
- Improved proxy service resilience for various response formats
- Better handling of OpenAI parameter restrictions in Codex mode

These changes provide a more reliable and specification-compliant implementation of the Codex chat completions endpoint while maintaining backward compatibility.

Author: CaddyGlow