Performance: `get_surface_distance` should use a KDTree for euclidian distance on CPU

[jberg5]

MONAI/monai/metrics/utils.py

Line 294 in eccefc5

dis = monai_distance_transform_edt((~seg_gt)[None, ...], sampling=spacing)[0] # type: ignore

Currently it does a full euclidian distance transform which computes all the distances between all of the points. On a GPU this is an embarrassingly parallel computation, but on a CPU we can get significant speedups from using a KDTree to just compute the distances to the nearest edge voxel.

I'll raise a PR - just creating this issue for tracking purposes.