ENH: allow start-stop array for indices in reduceat

[mhvk]

Triggered by #834 seeing some comments again, a draft just to see how it would look to allow reduceat to take a set of start, stop indices (treated as slices), to make the interface a bit more easily comprehensible without making a truly new method. It also allows passing in an initial to deal with empty slices.

Fixes #834

Draft only, since no new test cases yet. Mostly to discuss whether we want this at all, and, if so, what the API should be. So probably best not to worry too much about implementation (the duplication of code, both in reduceat itself and with reduce is large).

Sample use:

a = np.arange(12)
np.add.reduceat(a, ([1, 3, 5], [2, -1, 0]))
# array([ 1, 52,  0])
np.minimum.reduceat(a, ([1, 3, 5], [2, -1, 0]), initial=10)
# array([ 1,  3, 10])
np.minimum.reduceat(a, ([1, 3, 5], [2, -1, 0]))
# ValueError: empty slice encountered with reduceat operation for 'minimum', which does not have an identity. Specify 'initial'.

Writing it out like this, I think a different order may be useful, i.e., np.add(a, [(1, 2), (3, -1), (5, 0)]). The reason I picked the other one was that I liked the idea of triggering it by using slice(start, stop), with both start and stop possibly arrays and a tuple of two lists was closer to that (although internally it just turns it into an array). The list of tuples suggests more a structured array with start and stop (and step?) entries.

p.s. Fairly trivially extensible to start, stop, step.