Entity Recognition on sub-words

[tabergma]

Description of Problem:
Related to #5475

We found other edge cases that can happen if we are using a tokenizer that splits up words into sub-words. Let's take a look at an example:
Sentence: Buenos Aires is a city
Tokens: Buen, os, Ai, res, is, a, city

Scenario 1:
One entity covers multiple words or a single word.
city entity -> Buen os Ai res
type entity -> city

Scenario 2:
An entity covers just a part of a word.
city entity -> Buen

Scenario 3:
An entity covers two words, but at least on of the words just partly.
city entity -> os Ai

Scenario 4:
The sub-words of one word are annotated with different entities.
city entity -> Ai, state entity -> res

Scenario 1 and 4 are handled. We need to take care of Scenario 2 and 3.

Overview of the Solution:
We should keep labels if possible. Extend the entities to cover complete words instead of just parts of the words.