Thanks for tracking this down @outofculture ! This LGTM, merging.

Hello,
as noted here:
cupy/cupy#4693 (comment)
the "indexing bug" which made this workaround necessary seems to be solved in CuPy now.
Actually I had to revert this workaround back to its previous version to get the masking working properly with the 10.6.0 version of cupy.
Would it make sense to revert this PR and take the workaround out again?

Thanks tracking this down @NilsNoreyson !

@outofculture you good with reverting this?

I'm actually a little reluctant to outright do away w/ the old codepath, maybe we have an if statement depending on the version of cupy? After we get to 2 years since cupy 10.0 was released (what NEP-29 suggests we deprecate numpy versions) we then get rid of the old method?