Add Search::LexiconfreeRNNTTimesyncBeamSearch

[larissakl]

Adds a lexiconfree timesynchronous beam-search algorithm for standard (non-monotonic) Transducers. At each timestep, multiple non-blank labels can be predicted (the maximum number is controllable via a hyperparameter), a hypothesis is finished in the current timestep if it has emitted a blank label. In the inner loop of a timestep, first, all active inner hypotheses are extended with blank so they become outer hypotheses. Then, the inner hyps are extended by non-blank tokens and are pruned. If there are already more than max-beam-size outer hyps, all inner hyps which are worse than the worst of the max-beam-size best outer hyps are removed. If no inner hyps are left, the inner loop is stopped. At the end of a timestep, the outer hyps are pruned again based on their length-normalized score.

The implementation is based on PyTorch's RNNTBeamSearch

Major To-Dos:

• Integration of sub-scorers and intermediate pruning
• Correct handling of sentence-end
• Testing