Fix stop_gradient inconsistent API

[angeloskath]

The stop_gradient documentation states that the argument should be a list of
variables. The Theano implementation crashes if the argument is a list of
variables and the CNTK implementation crashes regardless but expects a list anyway (since it passes it to combine). The TensorFlow implementation expects one variable but in case a list is passed it doesn't crash as long as the shapes are all the same.

This commit handles both cases as can be expected. The following code illustrates the current Keras behaviour.

val = np.random.rand(10, 3)
backends = [KTH, KTF, KC]
as = [k.variable(val) for k in backends]
bs = [k.square(a) for k, a in zip(backends, as)]

a, b = as[1], bs[1]
KTF.stop_gradient(a)  # works
KTF.stop_gradient([a, b])  # works but returns a tensor?

a, b = as[0], bs[0]
KTH.stop_gradient(a)  # works
KTH.stop_gradient([a, b])  # crashes

a, b = as[2], bs[2]
KC.stop_gradient(a)  # crashes
KC.stop_gradient([a, b])  # crashes

This PR makes the behaviour consistent. When passed a single tensor a single tensor is returned. When passed a list or a tuple a list is returned.

Although returning a list and expecting a list would be more consistent with the previous documentation it wouldn't be consistent with the code's behaviour which could result in breaking others' code. This change now is backwards compatible but makes the behaviour consistent and documents it.