GH-457: PyTorch DataLoader

[alanakbik]

This PR refactors the way datasets are loaded in Flair. Instead of loading all data sets into memory as a List of Sentence, we now utilize PyTorch's Dataset and DataLoader methods and allow the user to choose whether or not to load a data set into memory. This allows us to scale training to very large datasets that do not fit into memory.

This PR also changes the syntax of how to load datasets.

Old way:

from flair.data_fetcher import NLPTaskDataFetcher, NLPTask
corpus = NLPTaskDataFetcher.load_corpus(NLPTask.UD_ENGLISH)

New way:

import flair.datasets
corpus = flair.datasets.UD_ENGLISH()

To use streaming loading, i.e. to not load into memory, you can pass the in_memory parameter:

import flair.datasets
corpus = flair.datasets.UD_ENGLISH(in_memory=False)

closes #458 and #457 #426

[kashif]

👍