fix: add fallback for torchaudio/torchcodec loading and support PyTorch 2.6+ security policy

[terrenceyeyang-code]

Fix: RTX 50-series Compatibility & PyTorch 2.6+ Security Policy Support

🔍 Problem Description

Tested on NVIDIA RTX 5080 with Python 3.12 and PyTorch 2.6.0, two critical issues were identified:

1. TorchCodec/Torchaudio OSError: torchaudio.load() fails on Blackwell architecture GPUs due to torchcodec shared library loading errors. This crashes the inference WebUI and dataset preprocessing.
2. PyTorch 2.6+ Checkpoint Loading: A change in PyTorch 2.6 (defaulting weights_only=True) prevents loading checkpoints containing pathlib.PosixPath objects, causing an UnpicklingError when resuming training.

🛠️ Proposed Solution

• Audio Loading Fallback: Implemented a try-except wrapper for torchaudio.load(). If the C++ backend fails, it automatically falls back to librosa (CPU-based) to ensure the pipeline remains functional.
• PyTorch 2.6 Security Alignment: Added pathlib.PosixPath to torch.serialization.add_safe_globals to allow safe unpickling of metadata during checkpoint restoration.
• Code Cleanup: All new logic includes English comments for better maintainability.

🧪 Test Environment

• GPU: NVIDIA RTX 5080
• OS: Docker / Linux
• Python/Torch: Python 3.12, PyTorch 2.6.0+cu124
• Verification: Confirmed that inference, dataset formatting (2-get-sv), and training resumption work seamlessly with this patch.

Addresses compatibility for the upcoming generation of GPUs and the latest PyTorch stable releases.