Incorrect punycode encoding and decoding for non-BMP codepoints

While working on IDNA 2008 implementation I discovered that the current punycode implementation does not correctly encode or decode non-BMP Unicode codepoints. For example U+102F7 COPTIC EPACT NUMBER FIVE HUNDRED is encoded to xn--ib9bk1k instead of correct xn--r97c, that is in turn decoded as incorrect U+02F7 MODIFIER LETTER LOW TILDE. This breaks IDNA (including version 2003) handling for domains with non-BMP characters.