Restructure Tokenizer and Splitter modules

[alanakbik]

When initializing a Sentence object, with use_tokenizer=True (set by default), it triggered an import of the SegtokTokenizer in the Sentence constructor. According to this thread, there is a cost of re-importing modules and since we often create large numbers of Sentence objects, this is unnecessary overhead.

This PR makes the import global. To do this, it splits up the flair.tokenization module into tokenization (for all tokenizers) and splitter (for all sentence splitters). This way, there is no longer an import of flair.data in flair.tokenization.

[alanakbik]

However, it doesn't really seem to make a difference in runtime. Tested with

import timeit
t = timeit.Timer(setup='from flair.data import Sentence',
                 stmt='Sentence("a a a a a a a a a a ")')
print(t.timeit())