Update Tree-/LexiconfreeTimesyncBeamSearch to perform intermediate pruning between multiple LabelScorers

[SimBe195]

Add the option to configure multiple label scorers with intermediate pruning after each one in LexiconfreeTimesyncBeamSearch and TreeTimesyncBeamSearch. This is for efficiency gains so that for example when searching with CTC + LSTM LM some pre-pruning can be done based on the CTC scores before doing the more expensive LSTM calculations.

Speech::ModelCombination has a new parameter num-label-scorers and then builds a list of LabelScorers of that size. The search algorithms then use this list and iteratively score extensions and apply intermediate pruning for each scorer in this list in the decodeStep. To this end, the max-beam-size and score-threshold config parameters were changed from ParameterInt and ParameterFloat to ParameterIntVector and ParameterFloatVector. For example with 3 LabelScorers the order of steps would be:

1. Build extensions list
2. Score each extension with scorer-1
3. Apply score-pruning with first score-threshold to extensions
4. Apply beam-pruning with first beam-threshold to extensions
5. Score each extension with scorer-2
6. Apply score-pruning with second score-threshold to extensions
7. Apply beam-pruning with second beam-threshold to extensions
8. Score each extension with scorer-3
9. Apply score-pruning with third score-threshold to extensions
10. Create full LabelHypotheses from extensions
11. Perform recombination of LabelHypotheses
12. Apply beam-pruning with third beam-threshold to LabelHypotheses

[curufinwe]

Looks OK to me. In mainly looked at the interfaces and parameter names and did not thoroughly look at the changes to the search algorithm. I only have 1 request for changes (add logging) and one clarification question.

[larissakl]

Maybe we should also adjust the documentation at the begin of the search algorithms. At the moment it says "Uses a LabelScorer to context initialization/extension and scoring.". I would maybe somehow mention that one can use multiple LabelScorers with intermediate pruning.