Dot op only supports matrix/vector args; other ndims must go through tensor.dot to ensure optimizations can be applied