feat: Add support for Blackwell GPU architecture (sm_120) and fix compatibility issues

[Sunmingyang1987]

Summary

This PR adds support for NVIDIA Blackwell GPU architecture (sm_120), such as RTX PRO 6000 Blackwell Max-Q, and fixes several compatibility issues when running on newer PyTorch/CUDA environments.

Changes

Code Changes

• musetalk/utils/preprocessing.py: Add monkey-patch for torch.load to handle PyTorch 2.6+ weights_only=True change, allowing legacy checkpoint (dwpose) loading
• musetalk/utils/face_detection/detection/sfd/sfd_detector.py: Move SFD model path to models/s3fd/ for unified model management
• scripts/inference.py: Fix --ffmpeg_path parameter being ignored when system ffmpeg is already available in PATH

Documentation (README.md)

• Added SFD model (models/s3fd/s3fd.pth) to model directory structure
• Added preprocessor_config.json note for whisper model
• Added comprehensive Troubleshooting section covering:
  • Blackwell GPU (sm_120) support: PyTorch upgrade + mmcv 2.1.0 installation
  • PyTorch 2.6+ checkpoint loading error
  • FFmpeg missing PNG encoder

Tested Environment

• GPU: NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition (sm_120)
• PyTorch: 2.9.0 + CUDA 12.8
• mmcv: 2.1.0 (compiled from source)
• mmdet: 3.1.0
• mmpose: 1.1.0