feat: GPU enumeration with VRAM reporting (NVIDIA, AMD, Windows)

[JamesNyeVRGuy]

GPU enumeration with VRAM reporting (NVIDIA, AMD, Windows)

Adds enumerate_gpus() to device_utils.py — a cross-platform function that detects all available GPUs and reports their VRAM. Useful for status display, VRAM gating before inference, and multi-GPU selection in batch scripts.

Currently device_utils.py can detect which device to use (CUDA/MPS/CPU) but can't tell you what GPUs are available or how much VRAM they have.

What does this change?

Adds to device_utils.py:

• GPUInfo dataclass: index, name, vram_total_gb, vram_free_gb
• enumerate_gpus() -> list[GPUInfo] with multi-backend detection:
  • NVIDIA: nvidia-smi CSV query (index, name, total/free memory)
  • AMD Linux: amd-smi (ROCm 6.0+) with live VRAM usage, falls back to rocm-smi (legacy)
  • AMD Windows: Registry lookup via winreg (fixes the Win32_VideoController.AdapterRAM uint32 overflow that reports >4GB GPUs as 0GB)
  • Universal fallback: torch.cuda.get_device_properties() for any GPU torch can see

No new dependencies — uses subprocess, json, winreg (stdlib). torch is only imported in the fallback path.

How was it tested?

• NVIDIA (Linux Docker + Windows native): nvidia-smi path returns correct GPU name and VRAM
• AMD (Linux ROCm): rocm-smi path returns correct VRAM for RX 7900 XTX
• AMD (Windows): registry path correctly reads 64-bit VRAM size for 16GB+ cards
• No GPU: returns empty list, no errors

Checklist

• uv run pytest passes
• uv run ruff check passes
• uv run ruff format --check passes