All input is now coerced to text before being normalized.
This changes nothing under Python 2 but it allows bytes
to be passed to the function without a TypeError under Python3
(bytes are assumed to be utf-8 encoded text).
Thanks to trac user vajrasky for the report.