GeezerVison

The punycode IncrementalEncoder and IncrementalDecoder ran the whole-string transform on each chunk and ignored final, so chunked input produced wrong output (e.g. b'a-b-' instead of b'ab-'). Punycode encodes the positions of non-ASCII characters relative to the whole string, so a chunk cannot be transformed on its own.

They now derive from BufferedIncrementalEncoder / BufferedIncrementalDecoder and buffer the input until final is true, as suggested in the issue. codecs.iterencode()/iterdecode() now give correct results, and getstate()/setstate()/reset() work through the buffered base classes.

Behaviour note: a caller that uses the incremental classes directly with a whole label but never passes final=True now gets empty output, since nothing is emitted until the input is known to be complete. That is the documented contract of incremental codecs, but it is a visible change for such code.

punycode stays in broken_unicode_with_stateful in test_codecs, because the StreamReader/StreamWriter classes still use the stateless codec per chunk; this PR only fixes the incremental classes.

Closes gh-160308.

Prepared with AI assistance, then reviewed and tested before submission.