Fix mask broadcasting by squeezing category masks to 2D

[HarishPermal]

Normalize category_mask from (H, W, 1) to (H, W) with squeeze(-1) and use [..., None] when blending, to keep shapes consistent with RGB images.

category_mask.numpy_view() returns a 3D array; stacking/broadcasting without squeezing can create invalid 4D shapes and OpenCV/NumPy errors.