test_histogram_multidim failing

[stress-tess]

This has debug info for issue #3366, since I'm unable to reproduce this locally

Apparently it's failing on the nightly machine even in single locale mode. This make me think this a numeric precision issue somewhere with float math. It's really strange because the problem size lands us in the + reduce case which has the simplest logic and very closely mirrors what is done in the 1d histogram case

we're seeing theses results for the first index of the 2d hist:

[11. 10. 12.  9. 10. 12.  7.  9. 11.  6.]
[10 10 11  9 10 11  7  9 11  6]

so we appear to undercounting by 1 inconsistently. I'm curious if these numbers are the same from run-to-run