fix: support RTX 50-series (Blackwell sm_120) CUDA via cu128 wheels (#217)

[thcp]

Problem

NVIDIA RTX 50-series cards (Blackwell, compute capability sm_120) fail GPU stem extraction with:

RuntimeError: CUDA error: no kernel image is available for execution on the device

The stock torch 2.6 cu124/cu121 wheels were built before Blackwell and contain no sm_120 kernels. torch.cuda.is_available() still returns True, so detection "succeeds" and demucs is told to use cuda — then dies the moment a kernel runs. Blackwell support requires the cu128 wheels, which only exist for torch 2.7+.

Fixes #217.

Approach — targeted Blackwell carve-out

Only sm_120 / sm_100 GPUs get torch 2.7.1+cu128; every existing user stays on the validated 2.6.0 path, so there is no regression risk for the current base. The 2.7 wheel is fetched at runtime only — pyproject.toml and the build scripts are untouched.

The torch<2.7 pin exists because torchaudio 2.7+ changed its codec backend, but stems are written via torchaudio.save() which StemDeck already routes through soundfile/libsndfile, independent of the torch version — so the 2.7 wheel writes WAVs fine.

Changes (all in desktop/src-tauri/src/main.rs)

• Compute-capability detection — detect_nvidia_gpu() now also runs nvidia-smi --query-gpu=compute_cap (new detect_compute_cap() helper). Missing/N/A is non-fatal and falls back to the old logic.
• wheel_tag() — routes compute-cap major >= 10 → cu128; everything else keeps the existing cu124/cu121/cu118 table.
• Version-aware install — install_cuda_torch() installs torch==2.7.1+cu128 for the cu128 tag, 2.6.0 for all others (unchanged).
• Hardened verification — verify_cuda_torch() now forces a real kernel launch instead of trusting is_available(), so an incompatible wheel falls back to CPU cleanly instead of crashing mid-job. This is a safety net independent of the cu128 fix.
• Unit test — wheel_tag_routes_blackwell_to_cu128 covers Blackwell/Ada/Ampere/Turing/missing-cap.

Verification

• ✅ cargo check compiles
• ✅ cargo test wheel_tag passes
• ✅ cargo fmt --check clean
• ⏳ Pending: validation on real RTX 5060 Ti hardware (@decrimtv-glitch offered to test) — confirm setup.log shows the cu128 install, cuda_verified = true, and a full GPU extraction with no error.

Note: 2.7.1+cu128 should be confirmed as the desired patch against the live PyTorch index before cutting the validation build.