Make the scores more tolerant for subword unit parts

[M4t1ss]

Penalize tokens ending with @@ less for having attention aligned to multiple other tokens.
...or maybe concatenate the attention matrix to word-level?