GH-564: regression datasets

[alanakbik]

This PR fixes a problem with the loss function in the (still beta) TextRegressor and adds 4 regression datasets.

You can now train a regression model like this:

# get the corpus
from flair.datasets import WASSA_SADNESS
corpus = WASSA_SADNESS()
print(corpus)

# choose some word embedding or stack of embeddings
embeddings = OneHotEmbeddings(corpus)

# choose some document embedding
document_embedding = DocumentPoolEmbeddings([embeddings], fine_tune_mode='none')

# init text regressor
classifier = TextRegressor(document_embedding)

# init model trainer
trainer = ModelTrainer(classifier, corpus)

trainer.train(
    f"resources/taggers/test-regression",
    mini_batch_size=4,
    max_epochs=20,
)

[alanakbik]

👍