Set more efficient default for mask type

[CodyCBakerPhD]

With mask_type='image' a relatively big segmentation dataset took forever and worse, stalled out in unstoppable ways (unless you completely kill the process from outside)

But swapping to pixel worked instantaneously. Proposing to set it as the default for both speed and reduced memory usage

[bendichter]

I suppose this is fine. Let's add an entry to the changelog though since this will change behavior for some users.

[CodyCBakerPhD]

Actually it could be a breaking change now that I think about it

image always work, but is the most inefficient for a large number of ROIs and/or large image resolution

pixels only works for 2D (non volumetric) imaging, so though it's more efficient specifying it in this way would break all volumetric pipelines that don't already override the option

Maybe a better distinction for the argument could be full or sparse (with sparse auto-detecting pixel or voxel appropriately based on image dimensions)