added rtx 5000 series support with torch 2.9 and cuda 13

[RaidOuahioune]

Added Modern GPU Support (PyTorch 2.1+ / CUDA 12-13)

Summary

Adds support for modern NVIDIA GPUs (RTX 30/40/50 series) by enabling TF32 acceleration and updating PyTorch compatibility.

Changes

app.py

• Enable TF32 for Ampere and newer GPUs (~3x speedup)
• Support both PyTorch 2.1-2.8 (old API) and 2.9+ (new API)
• Fully backward compatible

requirements.txt

• Update PyTorch to >=2.1.0 (broader compatibility)
• Update deepspeed to >=0.14.0 (CUDA 12+ support)
• Add torchcodec>=0.1.0 (required for PyTorch 2.8+)

Tested on:

• GPU: RTX 5070 Ti (Blackwell, cuda 13, Driver Version: 580.65.06)
• PyTorch: 2.9.0+cu128, CUDA 12.8