Always return the same data type

[chrisroat]

I noticed this in the recent release:

added support for >2^16 masks with np.uint32, if there are <2^16 masks then the masks are returned as np.uint16 still

This makes it extremely hard to use in a distributed setting like dask where they data types must be known in advance and should be the same for every chunk (you are effectively indicating that data chunks with fewer masks may be a different data type).

If you are internally using a bigger data type and changing it when #masks<2**16, then this is a pre-optimization that people may not need -- you can just leave that optimization to the user anyway.

It would be nice to use that same standard as scikit-image, and always return the same data type (preferable signed). This allows a user to "drop in" cellpose to their code without having to potentially rework around the API.