fix: use getLlamaGpuTypes("supported") to detect actual GPU availability

[DanieleSalatti]

Problem

On Linux machines without NVIDIA GPUs, QMD attempts to compile llama.cpp with CUDA support on every invocation. This adds ~30 seconds of noisy CMake failures to every qmd command before falling back to CPU.

Root cause

getLlamaGpuTypes() called without an argument returns all GPU types valid for the platform — on Linux that's always ["cuda", "vulkan", false] regardless of installed hardware. QMD then tries getLlama({gpu: "cuda"}), which triggers a source build of llama.cpp with GGML_CUDA=1, fails because there's no CUDA toolkit, and falls back to CPU.

Fix

Pass "supported" to getLlamaGpuTypes(), which checks for actual driver/library presence before including a GPU type. This is the intended API per node-llama-cpp docs.

Also removes the now-unnecessary @ts-expect-error directive — "supported" is a properly typed argument, so the type suppression was masking the issue.

Testing

Verified on Intel NUC (i7-1165G7, no discrete GPU) running Arch Linux

Before: every QMD command triggers ~30s of CUDA CMake build spam

After: instant startup, correctly detects CPU-only and skips GPU entirely

tsc passes clean with zero errors

[DanieleSalatti]

I failed at copy & pasting the draft, updating description