Don't use a separate function to handle "complex" attributes. Validate
UTF-8 byte sequences without decoding. This should improve performance
considerably when parsing multi-byte UTF-8 sequences.
Use a string buffer to avoid unnecessary allocations and copying when
expanding entities.
Normalize attribute values in a single pass while expanding entities.
Be more lenient in recovery mode.
If no entity substitution was requested, validate entities without
expanding. Fixes#596.
Also fixes#655.
