Inconsistent tokenization in Italian depending on the version of regex

[rspeer]

We tried to standardize the tokenization of French words such as "l'heure" across different versions of mrab's regex module. The fix assumed that we want this to come out as ["l'", 'heure'], and to recognize this pattern as one or two letters, an apostrophe, and a vowel.

A similar pattern appears in Italian, but it can have four characters before the apostrophe.

On regex 2020.4.4, we get this tokenization:

>>> wordfreq.tokenize("nell'obolo", 'it')
["nell'obolo"]

But on regex 2018.2.21, we get:

>>> wordfreq.tokenize("nell'obolo", 'it')
["nell", "obolo"]

This should be standardized as well. If we insist on a particular version of regex, we will probably cause conflicts with other libraries such as spacy.

[rspeer]

Closed now that we require regex 2021 or later.