Decrease Markov network memory footprint

[ForNeVeR]

I believe that currently Markov network images are consuming more memory that it's needed for them. We need to profile Horta memory and decrease network memory footprint if that's possible.

[ForNeVeR]

Please first of all check #134 when implementing.

[ForNeVeR]

Some optimization points (things to consider):

1. Intern strings (words) after parsing the file.
2. Intern phrase vectors inside Platonus library.
3. Make a single parser actor (so it won't eat too much memory while parsing multiple networks at once).
4. Progressively load strings from the database (don't load them all at once).