Memory management for training on large data sets

[alanakbik]

In uses cases where training data sets are large or there is little available RAM, language model embeddings cannot be stored in memory (see #135 ).

Current solution: Currently, the only way to still train a model in such cases is to set the
embeddings_in_memory tag to False in the trainer classes (TextClassifierTrainer or SequenceTaggerTrainer). With this flag, embeddings get generated on the fly at each epoch and immediately discarded after use, which solves the memory issue but is computationally expensive since there is no re-use of already computed embeddings.

Idea: Use a key-value store to persist embeddings to disk and enable quick lookup of already computed embeddings. A nice side-effect is that if we run several experiments on the same dataset, embeddings from earlier runs can be re-used, thus speeding up parameter-sweep experiments.

[alanakbik]

will be part of release-0.3 and activated by default for CharLMEmbeddings (can still be turned off to save disk space)