Memory requirements

[jlibovicky]

I tried to evaluate all Flores languages and ran out of memory on a machine with 40GB memory. Can we make it somehow less memory demanding?

[varisd]

Currently, we are reading all the provided datasets into memory because they are accessed multiple times (at minimum, by each metric), the current implementation would otherwise require reading the files multiple times costing time.

We can probably:

1. add cacheing to the TokEvalData (keeping only limited amount of text in the memory and loading the dataset only when required) and
2. rethink how TokEvalScorer iterates over the data (this might be at the cost of losing expresiveness of the TokEvalMetric though).

I agree that the current implementation is very brute-force.