Feature: automatic GPU variant selection (CUDA + Vulkan) in loadBinary

[lloyal-research]

Summary

loadBinary() selects a GPU variant only from LLOYAL_GPU / the gpuVariant arg; no consumer (reasoning.run, rig) sets it, so an installed GPU silently runs on the CPU prebuilt. We want loadBinary() to auto-select the best available prebuilt when the caller hasn't asked for one.

A CUDA-only first cut was prototyped in PR #32 and pulled out (it was incomplete and out of scope for the #20 crash fix). This issue tracks doing it properly. (Note: once the #20 AVX-512 fix lands, LLOYAL_GPU=cuda already works again — auto-detect is a convenience feature, not part of the regression fix.)

Scope — all supported variants

From the release matrix:

• Metal — no work needed: there is no -metal package; darwin-arm64 IS the Metal build (llama.cpp auto-enables it). The base package already accelerates.
• CUDA — NVIDIA. Detect the driver: libcuda.so.1 (Linux, incl. WSL2 /usr/lib/wsl/lib/), nvcuda.dll (Windows), with an absolute-path nvidia-smi fallback. No PATH lookup (hijack-safe).
• Vulkan — AMD/Intel/Qualcomm (and NVIDIA). Detecting the loader (libvulkan.so.1 / vulkan-1.dll) is not enough — a software/llvmpipe device would false-positive. Must enumerate a real hardware physical device (e.g. a tiny probe or parsing vulkaninfo).
• CPU — fallback.

Design requirements

1. Preference order: a box can have both CUDA and Vulkan. Prefer CUDA on NVIDIA; else Vulkan if a hardware device exists; else CPU. Document and make overridable.
2. LLOYAL_NO_FALLBACK contract (important): it must apply only to an explicitly requested variant (LLOYAL_GPU / gpuVariant). An auto-detected variant must always fall back to CPU if its package fails — otherwise auto-detect turns a working CPU load into a throw for users who set NO_FALLBACK for unrelated reasons. (This was the high-severity bug in the PR fix: pin AVX2 ISA floor for x64 prebuilts (hdk#20) #32 prototype.)
3. stdout stays clean — diagnostics to stderr (console.warn), cached probe, never throws.
4. CI invariant: every CI job that loads the binary sets LLOYAL_GPU (or LLOYAL_LOCAL), so auto-detect is end-user-only and never depends on CI-provisioned paths. Keep it that way; add a test asserting auto-detect is skipped when a variant is explicit.
5. Reliability: containers, WSL2, no-GPU servers, non-FHS installs. Best-effort with the LLOYAL_GPU escape hatch — a miss must be non-fatal (CPU).

Reference

Prototype commits (reverted from PR #32): the CUDA driver/nvidia-smi probe and the loadBinary integration — reuse as a starting point for the CUDA half, then add Vulkan + preference + the NO_FALLBACK fix.