Distinction between rank-0 arrays and scalars

[asmeurer]

There are some instances where reduce currently transforms one index into another where they don't give the same thing on scalars or rank-0 arrays. For example

>>> np.int64(0)[...]
array(0)
>>> np.int64(0)[()]
0
>>> np.array(0)[...]
array(0)
>>> np.int64(0)[...]
array(0)

Currently ellipsis().reduce() gives Tuple().

>>> np.array([1, 2, 3])[0,]
1
>>> np.array([1, 2, 3])[0,...]
array(1)

Currently the code assumes that tuples always implicitly end in an ellipsis (Tuple.reduce removes any trailing ellipsis) .

I don't know if we should care about this distinction.

[asmeurer]

This will happen in more cases after fixing #31. For now, I am adding a note to Tuple.reduce about it indicating that ndindex does not distinguish between rank-0 arrays and scalars.

[asmeurer]

We could fix this by adding an allow_rank0_equals_scalar flag to the reduce methods. If False, indices won't reduce to each other if one can produce a scalar but the other a rank-0 array. The default should likely stay as True, since False would prevent removing redundant ellipses in many cases, and for most use-cases, scalars and rank-0 arrays are interchangable.

I don't plan to implement this soon, as I don't believe I will need it.