feat(gpu): offer the CUDA pack to Pascal cards (0.0.9 kernel floor)

[gabrielste1n]

⏸ HOLD — do not merge until the Pascal smoke test below passes on real hardware.

With #1584 pinning the packs to whisper.cpp 0.0.9 (which ships sm_61 kernels), this lowers MIN_CUDA_COMPUTE_CAP from 7.5 to 6.1 so GTX 10-series cards are offered the CUDA pack.

Exactly who this affects

• Cards ≥ 7.5, AMD/Intel, macOS, cloud users: no change — they already passed the gate or never consult it.
• Cards < 6.1 (Maxwell/Kepler): no change — still routed to Vulkan (new below-floor test added).
• Pascal/Volta (6.1–7.4): newly offered the CUDA pack. Volta runs via the embedded compute_61 PTX (driver JIT).

Prefer-installed-backend fix included

Without it, a Pascal user already running the Vulkan pack would see the "Enable GPU" CUDA banner reappear over a working setup the moment this gate opens. The picker now prefers whichever pack is installed — mirroring the runtime resolver, which only prefers CUDA once it's actually downloaded. Their dictation never changes out from under them either way.

Bounded worst case (already shipped)

If a Pascal kernel still fails on some card: the dictation completes via the CPU fallback retry, the failure is persisted (WHISPER_GPU_FAILED) so it's never re-attempted silently, and the chip shows "GPU could not be activated — Retry". Nobody loses words; the cost is one wasted download.

Go / no-go: Pascal smoke test

On a GTX 10-series machine running this build:

1. Settings → Speech-to-Text → Local → Whisper → Enable GPU (should offer CUDA, ~772 MB).
2. Wait for the chip to reach "GPU acceleration active" (poll reflects the live server, not the download).
3. Dictate a long sentence; confirm it pastes, no "using CPU instead" toast appears, and whisper-server-status reports gpuBackend: "cuda".
4. Compare latency against the Vulkan pack on the same box (both should beat CPU; CUDA should beat or match Vulkan).

Full suite: 1923 pass / 0 fail, typecheck + prettier clean.