Improve recognition parameter handling

[christophmluscher]

When doing grid search on the decoding parameters for a hybrid model, doing such a cartesian product does not make so much sense. One should first tune model-related scales such as prior scale and tdp scale. Then tdp values, and only given the optimal values of the mentioned parameters one tunes the lm and pronunciation scales. Moreover, we should definitely consider the obligatory use of a high altas and small beam for the first two steps - only lm scale should not be tuned together with altas.

Please also consider that experience shows that the only tdp value that is worth it to tune is the exit penalty for silence and non-word.

Originally posted by @Marvin84 in #110 (comment)

[Marvin84]

The idea is to have three steps, implemented as independent functions of the HybridDecoder class.

Set a small beam (e.g. 14, and high altas, e.g. 12) and run the next 2 steps

1. Tune the tdp and prior scales
2. With the optimal value of this tune exit penalties of silence and non-word

Set the final beam (e.g. 16 or 18, and lower altas, e.g. 2.0, 4.0) and then
3. Tune the LM scale
4. Optionally with all optimal values, tune the search space size. (we pruning, beam limit, beam, altas)