fix: disable VMM on WSL2 to prevent cuMemAddressReserve crash

[wzgrx]

Problem

When running on WSL2 (especially with CUDA), node-llama-cpp compilation or runtime often fails with a hard crash:

CUDA error: out of memory
cuMemAddressReserve failure

This is documented in Issue #580.

The root cause is that WSL2/WDDM drivers cannot allocate the 32GB virtual memory pool that llama.cpp attempts to reserve by default. Even if physical VRAM is available, the virtual address reservation fails, leading to an abort.

Solution

This PR automatically detects WSL2 environments and injects -DGGML_CUDA_NO_VMM=ON into the CMake build options when compiling for CUDA.

Changes

1. WSL2 Detection: Added wsl2 flag to getPlatformInfo() by checking /proc/sys/kernel/osrelease for microsoft or wsl substrings.
2. Automatic VMM Disabling: Modified BuildCommand.ts to set GGML_CUDA_NO_VMM=ON when gpu=cuda and platformInfo.wsl2 is true.

Testing

• Tested on WSL2 (Ubuntu 24.04, RTX 5090, CUDA 13.2).
• Embedding tasks now complete successfully without OOM crashes.
• GPU offloading remains fully functional.

Fixes #580