gh-154859: Keep the iconv shift state across incremental decode calls

[fedonman]

The iconv codecs opened a fresh conversion on every call, so decoding a stateful encoding in
chunks lost the shift state and silently returned wrong text:

one-shot:    ABC中文DEF
incremental: ABCVPNDDEF

_codecs.iconv_state() now opens a conversion that the incremental decoder and the stream
reader hold and pass to iconv_decode(), so one conversion spans the stream. reset()
starts a new one. The capsule closes the conversion when it is collected.

Only decoding is affected. Incremental encoding re-emits the escape sequences per chunk,
which is more verbose than the one-shot output but decodes back to the same text.

The tests fail without the change. Note that encodings._iconv_codecs is frozen, so a
rebuild is needed to see that. Checked with 50k decoders for a leaked conversion: no RSS
growth. This build has no --with-pydebug, so I could not run -R 3:3.

The iconv codecs are new in 3.16, so there is no NEWS entry.

• Issue: iconv codecs corrupt stateful encodings when decoding incrementally #154859

[BHUVANSH855]

These regression tests currently construct the test input with:

data = codecs.encode(text, 'iconv:' + enc)

From the iOS CI failures, it looks like iconv:ISO-2022-CN is available but cannot encode the sample text on that platform, so codecs.encode() raises UnicodeEncodeError before the incremental decoding path is exercised.

Would it make sense to either skip the test when the sample text cannot be encoded on the current platform, or make the test independent of platform-specific iconv encoding support (for example by using precomputed encoded bytes)?