Implemented gradient of CrossentropySoftmaxArgmax1HotWithBias when there is an incoming gradient on the softmax's output

for the duration of a single function evaluation, rather than globally at the
creation of a DebugMode function

Move the generation of indices used for iterating over broadcasted tensors to _generate_broadcasting_indices

Added more explicit error message when trying to use verify_grad on something with more than one output

Fixed recent buggy implementation of CrossentropySoftmax1HotWithBiasDx.grad and added unit test to check it