Vocabulary optimization for multithreading speedup

[fhieber]

• removed MurmurHash and replaced with HashMap
• lazy loading of getNonTerminalIndices
• use of faster FormatUtils.isNonTerminal()
• Consistent use of locks.

This change fixes a rare multithreading race condition issue when an empty vocabulary (in the case of no grammars loaded and lots of input data is passed to the decoder) was filled by multiple threads. In addition, by storing the list of nonTerminalIndices in the vocabulary the iteration over all known words for every input sentence is avoided. This provides a significant speedup in multithreaded decoding. MurmurHash functionality was removed for code simplicity.

[mjpost]

Awesome, I am testing this now.

[mjpost]

I tested on a Europarl phrase-based grammar, with very nice results (6-CPU machine, with 12 using hyperthreading):

I'll now just run it past the regression tests and the merge it in.