Fix various bugs in np.ma.where - #8647

Sorry about that, unexpectedly caused some tests to fail when trying to clean something up in the rebase

For subclasses, one now has the problem as np.where itself doesn't do subclasses correctly

Although that would then lead to confusion when np.where(masked_arrays) returns a masked array that doesn't have the correct mask.

So if np.where supporting subclasses is risky, perhaps supporting subclasses here is also the wrong thing to do, since it's hard to conceive of a useful subclass that would keep information without having a chance to __array_wrap__

Do you want me to revert that last commit then?

@eric-wieser - I have no strong feelings about this. Overall, a well-written class with a good __array_finalize__ just might work OK with your new construct (given a working np.where), and I don't really see any downside, so let's just stick with what you have. If you agree, then I think this is ready to go in (but if you prefer just to do the first commit, that is fine too)

I think I can be convinced to keep this, in the case of something like a MaskedQuantity subclass that just needs to copy metadata?

@mhvk: I rolled back that last commit - I think np.where not supporting subclasses is a strong argument that np.ma.where shouldn't either.

@mhvk: Tests all pass - good to merge?

Great! Merging...