Local tests pass. Ready for review @seberg

This is that last thing that would be great to have some fix for in 1.24.1 considering that it returns bad results. I expect this is good, but I need fresher eyes. But if it looks good to others: Maybe we should get it in, and I look at it again later and just follow up if I feel a different approach is better.

I'm thinking about this proposed change now and while it does fix the problem, it is more conservative than necessary. Recall we are looking for overflows in this calculation:

basic_mask = (ar1 <= ar2_max) & (ar1 >= ar2_min)
outgoing_array[basic_mask] = isin_helper_ar[ar1[basic_mask] - ar2_min]

basic_mask will trim ar1 to only the elements within the range of ar2. Thus, we only technically need to consider min(ar1_max, ar2_max), and max(ar1_min, ar2_min), in these calculations. i.e., the following:

# After masking, the range of ar1 is guaranteed to be
# within the range of ar2:
ar1_upper = min(int(ar1_max), int(ar2_max))
ar1_lower = max(int(ar1_min), int(ar2_min))

range_safe_from_overflow &= all((
    ar1_upper - int(ar2_min) <= np.iinfo(ar1.dtype).max,
    ar1_lower - int(ar2_min) >= np.iinfo(ar1.dtype).min
))

does that make sense?

Edit: pushed this change.

I wonder why integers larger the uint16 are not tested, are they too big?

Thanks @MilesCranmer. @seberg If you see anything that bothers you we can make another PR.