[compiler] fix ndarray broadcasting with 0 sizes

[patrick-schultz]

Really bad example:

In [2]: a = hl.nd.array(2)

In [3]: empty = np.array([], np.int32).reshape((0, 2, 2))

In [4]: empty = hl.nd.array(empty)

In [5]: hl.eval(empty)
Out[5]: array([], shape=(0, 2, 2), dtype=int32)

In [6]: hl.eval(a + empty)
Out[6]:
array([[[       2, 55622722],
        [       2,        3]]], dtype=int32)

Result of sum of shape () with shape (0, 2, 2) should be shape (0, 2, 2) by standard numpy broadcasting rules. Instead, we try to produce a result with shape (1, 2, 2), which contains garbage data.

[ehigham]

looks great. glad to see tests enabled!

[danking]

WIP until I sort the release.