Grad check fails for matrices where the max is close to the second highest value

[AngledLuffa]

I guess this is to be expected, actually, but if the difference between the max in a region and the next highest value is less than delta, the next highest value will have a numerically calculated gradient > 0, causing the grad check to fail.

[AngledLuffa]

This probably isn't worth fixing, actually, but it comes up if you repeat the random gradient check enough times.