Corpus and Span label refactoring

[alanakbik]

First part of a larger refactoring of corpora and span logic in Flair. The refactoring is motivated by difficulties in scaling the EntityLinker up, which required looking at Span representations and the corpora.

1 Corpora
The PR unifies several corpora into a single object. Before, we had ColumnCorpus, UniversalDependenciesCorpus, CoNNLuCorpus, and EntityLinkingCorpus, which resulted in too much redundancy. Now, there is only the ColumnCorpus for all such datasets.

This makes a number of changes:

• You can now specify a min_count when computing the label dictionary. Labels below that count will be UNK'ed. (e.g. tag_dictionary = corpus.make_label_dictionary("ner", min_count=10))
• The Dictionary will now compute count statistics for labels in a corpus
• The ColumnCorpus can now handle relation annotation, dependency tree information and UD feats and misc

2 Span Labels
We now make special distinction between token-level and span-level labels. Instead of storing entities as BIOES tags, they are now stored as span-level annotations. The SequenceTagger internally converts these to BIOES/BIO tags during training. At prediction time, BIOES tags are interpreted and span-labels added.

This makes a number of changes:

• You now choose the labeling format when instantiating the SequenceTagger, i.e.

    tagger = SequenceTagger(
        hidden_size=256,
        embeddings=embeddings,
        tag_dictionary=tag_dictionary,
        tag_type="ner",
        tag_format="BIOES",
    )

• the get_spans() method of the Sentence is removed for now, a similar method will be added back in when span-refactoring is complete
• The ColumnCorpus will automatically identify which columns are span labels and treat them accordingly

The PR also makes a number of other changes:

• EntityLinker class refactored for speed
• performance improvements in standard evaluate() method, especially for large datasets
• A new WordTagger class is created for simple word-level predictions
• ColumnCorpus no longer does disk reads when in_memory=False, it simply stores the raw data in memory leading to significant speed-ups on large datasets
• The deprecated data_fetcher is finally removed
• the Multiconer corpus object was renamed to NER_MULTI_CONER to match names of other corpora