einsum optimization may change dtype of output

[toslunar]

intermediate array may have ndim=0.

>>> x = np.arange(3).astype(np.int64)
>>> y = np.arange(2).astype(np.int16)
>>> np.einsum('i,i,j', x, x, y, optimize=['einsum_path', (0, 1), (0, 1)])
array([0, 5], dtype=int16)
>>> np.einsum('i,i,j', x, x, y, optimize=['einsum_path', (0, 2), (0, 1)])
array([0, 5])

[eric-wieser]

As I think you might have already worked out, this is caused by #10322