When decoding an IDN label that doesn't contain any ASCII 
characters, the function puny-decode-string-internal would skip 
the first letter, thereby returning gibberish instead of the 
expected result.

That is, the example in the puny-decode-string docstring, 
"xn--bcher-kva", would be decoded correctly since it finds the 
"kva" part after the last dash, but my domain "xn--9dbdkw.se" 
would not, since there is no "last dash".

This patch seems to fix the problem: