Make thought through decisions on string normalization defaults

[yngvem]

• Should we use NFC or NFKC?
• Should we convert all spaces to the same type?

[yngvem]

Do we want tokenizers to strip? Maybe not, because if we tokenize a token, then it should probably not change, but stripping before tokenizing breaks this property.

Consider grapheme cluster tokenizer: 'a a' --- tokenize ---> ['a', ' ', 'a'] --- tokenize tokens ---> ['a', '', 'a']