CI never runs pytest; the GPU job builds a multi-GB PyTorch image just to lint; the GPU code is never exercised

[joeljose]

Severity: High. Regressions like #13-#16 can merge without being noticed.

Problems

1. Tests don't run in CI. Both jobs run lint, import, --help and --version. The 28 CPU tests (27 s locally) run only through ./test.sh.
2. The GPU job is heavy for what it checks. test-gpu builds on pytorch/pytorch:2.1.2-cuda12.1-cudnn8-runtime (several GB) and installs pytorch_wavelets from git on every push and PR, only to repeat the same lint, import and help checks.
3. The GPU code never runs anywhere automated. tests/test_visualmic_gpu.py skips without CUDA.

Suggested fix

• Add python -m pytest tests -q to the CPU job.
• Run the GPU code path on the CPU in CI. pytorch_wavelets runs fine on CPU tensors. During this audit, extract_audio_gpu ran end to end on a CPU-only PyTorch build (torch 2.14 CPU wheel with pytorch_wavelets from git) after patching torch.device('cuda') to CPU. That run found GPU path drops the last partial batch when the container over-reports frames; CPU path drops frames when it under-reports #16 and showed CPU vs GPU-path correlation of 1.000. Make the device a parameter (--device cpu|cuda:N, see the CLI issue) so tests don't need monkeypatching.
• Replace the heavy test-gpu image build with a CPU-torch job (pip install torch --index-url https://download.pytorch.org/whl/cpu), and build Dockerfile.gpu only when it or requirements-gpu.txt change (paths: filter).
• Add an end-to-end CLI test on a tiny synthetic clip, generated in a fixture, that checks the WAV exists with the expected sample rate and length.
• Add permissions: contents: read to the workflow.

Acceptance criteria

• CI fails on any failing test.
• The GPU code path runs in CI on the CPU.
• The heavy GPU image isn't rebuilt on every PR.