Avoid operations on uninitialised memory in HMMs

[MBradbury]

Fix for #1078 by using torch.full to initalise tensors full of -inf instead of using torch.empty and then subtracting inf. Otherwise there is the potential for these tensors to contain NaNs if the uninitalised memory contains values such as inf or -inf.

The tensors initalised with torch.zeros are correct, but have been changed to match the other tensor initialisation.