BUG: downcasting from complex in `apply_unitary`

The following code to multiply PauliY by the identity incorrectly casts the matrix representation of PauliY into the type of the input state (float):

import cirq
import numpy as np
from cirq import protocols

y0 = cirq.PauliString({cirq.LineQubit(0): cirq.Y})
state = 0.5 * np.eye(2)
args = protocols.ApplyUnitaryArgs(target_tensor=state,
                                  available_buffer=np.zeros_like(state),
                                  axes=(0,))
result = protocols.apply_unitary(y0, args)

gives

>>> print(result)
[[0. 0.] 
[0. 0.]]

numpy correctly throws a warning about this casting, but this should really throw an error.

[dabacon]

This feels like a case where the code should use the larger type.

[tanujkhattar]

From cirq sync: We should throw an error.