Segment long text automatically and manage sliding context

[kermitt2]

• Long text are segmented automatically on server side given a target size (default 1000 characters, ~150-200 words), using in priorities end-of-line (double then single), then paragraph boundaries. Segmentation is balanced to have segments of similar sizes.
• DocumentContext object is created after processing of the first segment and refreshed sequentially when processing the next segment

Advantages:

• faster with large texts
• better results because models are based on a paragraph size
• more stable results and predictions independently of the size of the input text (solves Impact of text length on identified entities #131)
• remove the burden of segmenting text and managing a context on the client side

Tests added for the text segmentation. Default target segment size is 1000 characters, but this is also be configurable in the query. Documentation updated. Support the edge/pathological case of sentences of more than 1000 characters