Fix memory management for Flictionaries & prediction algorithm

[patrickgold]

This PR is a response to the reported memory crashes in #403 and attempts to fix them. In detail, this is what has been improved:

• The Flictionary load function previously created a lot of unnecessary allocations which it does not need. Additionally, the input stream of the binary raw file is now read as a stream, instead of per-caching the whole file before working with it.
• The NgramNode representation has also changed, so each node is much smaller in memory size than before. Also, the child nodes are now stored in an ArrayList rather than a LinkedHashMap, which vastly reduces the allocated memory size.