Sum on 3 dimensions raises NotImplementedError

[CharlelieLrt]

This might be already documented elsewhere, but when trying to execute a numpy-based code with cunumeric in place of numpy, some of the reductions such arr.sum(axis=(-3, -2, -1)) fails with NotImplementedError: Need support for reducing multiple dimensions.

A simple workaround is to instead do it dimensions by dimension with arr.sum(axis=-1).sum(axis=-1).sum(axis=-1), but it would be nice to have the multi-dimensional reduction. Having this limitation mentioned in the documentation of cunumeric.sum could also be useful.

[manopapad]

Also requested by SLAC (look for "nansum not reducing over multiple dimensions" at #1116), but they have a workaround.

Also requested by @bittuh at #1195.

[manopapad]

Assigning to @bryevdv's to do a partial implementation, namely support multi-axis unary reductions by reducing one axis at a time. A "speed of light" solution should really wait until we have support for code-generating operations (instantiating templates for all combinations of dimensions would result in a combinatorial explosion and excessive bloat in our binaries).