fix: memory corruption and serialization issues (v0.5.21)

[justrach]

Summary

This PR fixes several critical issues discovered when running TurboAPI with free-threaded Python 3.13t and Metal/MLX GPU frameworks:

• Memory corruption in async handlers - Body bytes were being corrupted between Rust and Python due to improper memory ownership transfer
• Response object serialization - Response objects were serialized as their string representation instead of actual content
• Async handler parameter passing - Async handlers with BaseModel params received kwargs instead of the model instance
• JSON parsing with special characters - simd-json rejected valid JSON that Python's json module accepts

Changes

Core Fixes

• Use PyBytes::new() instead of as_slice() for proper memory ownership in async handlers
• Add defensive bytes(bytearray(body)) copy in Python to prevent corruption with Metal/MLX
• Route async handlers with BaseModel params through enhanced path for proper model instantiation
• Add Python json.loads fallback when simd-json fails

Response Handling

• Improve Response object detection via model_dump() and attribute checks
• Handle binary data (audio/wav) without UnicodeDecodeError
• Add base64 encoding fallback for binary content

Housekeeping

• Clean up .gitignore (remove duplicates, improve patterns)
• Add unit tests for all fixes
• Add uv.lock for reproducible builds

Test plan

• Unit tests pass for Response serialization
• Unit tests pass for async handler classification
• Unit tests pass for JSON parsing with special characters
• Manual testing with MLX voice generation server (async handlers with BaseModel params)
• Both sync and async handlers work correctly with Metal GPU loaded