test_orientation is failing with new Numpy's pre-release

[MarcCote]

In the next Numpy's release, boolean index must match indexed array along specified dimension.
See https://travis-ci.org/nipy/nibabel/jobs/186167556#L1230

From https://github.com/nipy/nibabel/blob/master/nibabel/orientations.py#L71

    P, S, Qs = npl.svd(RS)
    # Threshold the singular values to determine the rank.
    if tol is None:
        tol = S.max() * max(RS.shape) * np.finfo(S.dtype).eps
    keep = (S > tol)
    R = np.dot(P[:, keep], Qs[keep])

When RS is not square, keep will have a dimension less than either P's columns or Qs' rows, therefore indexing using keep will raise an error. I haven't examine the code but maybe we could use full_matrices=False when calling npl.svd() (see https://docs.scipy.org/doc/numpy/reference/generated/numpy.linalg.svd.html)