Improve prediction wall-time

[josephbirkner]

Currently, the performance of the completer-lstm drops approx. linearly with the length of the prefix string. This problem could be circumnavigated by caching the last LSTM-State-Tuple-List on the client side, and feeding it to the server when an extended completion is requested.

Furthermore, a significant share in walltime is occupied by JSON Serialization/Transmission/Deserialization of the M*N*(char, probability) prediction matrix, where N is the number of chars to predict and M is the total number of lexical features. To reduce the size of matrix to significant entries, M should only cover a top portion of the (5?) most probable next characters.