Add word dropout

[alanakbik]

In a recent guest talk at Zalando Research, @hanxiao made a strong case for using word dropout in text classification. So, let's add it to Flair and evaluate it vis-a-vis standard and locked dropout.

• Implement word dropout
• Evaluate in downstream tasks

[alanakbik]

First experiments are looking promising. With a word dropout of around 0.1 (i.e. approx. every tenth word gets dropped) we get at least equal and sometimes better results than without. We even saw a new SotA number for NER, but have to confirm with more experiments.

This will probably be included as feature in release 0.3