Change normalize_encodings() to avoid using .translate() or depending on

the string type.  It will always return a Unicode string.  The algoritm's
specification is unchanged.