Parallel acoustic composition, layer freezing configuration parameter

[m-toman]

This PR adds:

Parallel acoustic composition.
Test run with 100 sentences, 8 processes:

serial
2018-02-14 12:56:49,685 ^[[1;32mINFO    ^[[0m       main    : creating acoustic (output) features
2018-02-14 12:58:39,720 INFO           main    : normalising acoustic (output) features using method MVN

parallel
2018-02-14 12:45:05,138 ^[[1;32mINFO    ^[[0m       main    : creating acoustic (output) features
2018-02-14 12:45:32,097 INFO           main    : normalising acoustic (output) features using method MVN

Layer freezing configuration parameter

New parameter freeze_layers to specify the number of layers to freeze when doing transfer learning/speaker adaptation. This just adds the configuration parameter and uses the existing mechanism in deep_rnn.py.

[ronanki]

Thank you for the pull request.