Add power to each backend and as a Tensor method in linalg

[alewis]

We should support the numpy function np.power.

Edit: it would also be good to support elementwise exponentiation of Tensor using the ** operator