Fix CI: pin torch 2.10.0 and use torchcodec 0.10.0 from PyTorch CPU index

[Copilot]

torchcodec 0.11.0 from PyPI ships CUDA-linked .so files that dlopen libnppicc.so.13 (CUDA NPP) at import time — a system library absent on GitHub Actions runners, causing all test collection to fail.

Changes

• Pin torch ecosystem to 2.10.0: torch==2.10.0, torchaudio==2.10.0, torchvision==0.25.0, torchcodec==0.10.0 — at this version, the torchcodec CPU wheel has zero CUDA shared-library dependencies (only FFmpeg + torch)
• Route torchcodec through PyTorch CPU index: adds [[tool.uv.index]] (explicit) + [tool.uv.sources] so uv always resolves torchcodec from https://download.pytorch.org/whl/cpu, not PyPI, regardless of machine CUDA availability
• Delete uv.lock: stale lock pinned the CUDA PyPI wheel; CI regenerates it under UV_TORCH_BACKEND: cpu with the correct CPU-only resolution
• Restore torchcodec usage: reverts the interim torchaudio workaround — AudioDecoder is back in transforms.py, datasets/audio.py, and test_transforms.py

[[tool.uv.index]]
name = "pytorch-cpu"
url = "https://download.pytorch.org/whl/cpu"
explicit = true

[tool.uv.sources]
torchcodec = { index = "pytorch-cpu" }