The detection was failing with 'index out of bounds' when models output
single-channel or channels-first tensor formats. Now handles:
- Grayscale (H, W) -> expands to 3 channels
- Single channel (H, W, 1) -> replicates to 3 channels
- Channels-first (C, H, W) -> transposes to channels-last
- 2-channel images -> adds third channel
- >3 channels (e.g., RGBA) -> takes first 3

Also removed incorrect 8GB RAM check that was blocking detection.