perf: improve speed of 26, 4, and 8 connected continuous labeling

[william-silversmith]

Take advantage of structure in the continuous case.

• If two values match within a neighborhood, their mutual neighbors will be affected identically.
• If the "catbird" neighbor and the current voxel's combined "range" encompasses the range of the dataset, then the local area can be treated as a binary problem.
  • If this is true globally, then the entire problem can be treated with a binary decision tree (the one that has the most literature written about it). However, I don't think I have the time to do that right now.

[william-silversmith]

I think there's some more structure in there too to exploit, but that's more complicated. Let's just get this simple change out. It helps a lot on more uniform images, but doesn't really do much on more natural images.