Detect unencodable characters in attribute names (from namespace
prefixes) in XmlWriter.

Would it be better to have a different representation for prefixes?  Perhaps
a schema document level map with overrides.  What about when annotations get
moved between documents?
