Fixes prediction across all architectures

[kylebgorman]

Two bugs found by testing prediction across all architectures:

• Hard attention, some RNN-backer pointer-generator, and transducer models all inherit the beam search implementation in RNNModel, but this incompatible so it needs to be disabled. Failing to do this results in an inscrutable error instead of a NotImplementedError. This is fixed in all locations. In each case I made the decision to place the exception relatively "low" (i.e., more derived) in the class hierarchy, so if someone actually did implement beam_decode on these classes everything would just work.
• The base model implementation of the prediction loop is made to have what seems like the obvious polymorphic return type which restores compatibility for hard attention. I suspect this is a minor regression introduced in Added beam search for LSTM #257; it has been fixed.

Clean-ups done at the same time:

• Standardizes the name of the special symbol: it's called EOS at various points in the documentation but the code actually calls it END and its tag is <E>.
• Uses the names beam_decode and greedy_decode everywhere.

[Adamits]

Left one nit, but LGTM.

Thanks for taking the time to do this, you're doing the lord's work.