For some reason, the normal (and faster than oneshot) reduction algorithm does not obtain the same results as NumPy when computing the reductions all and any. Thus, right now we always use the oneshot algorithm in order to have the same behavior as NumPy. This should be fixed.