Pack mixed-width UTF-32 blocks on vector and fallback paths - #1018

Summary

The UTF-32→UTF-8 converters fall back to scalar code as soon as a block contains mixed widths. This PR keeps such blocks on the vector path:

• Haswell (AVX2): three comparison masks classify each lane and form two bits per code unit (its UTF-8 width); each group of four lanes is compacted with a PSHUFB plan indexed by those widths, packing eight valid code units at a time (new width-indexed tables).
• Fallback: validates a batch plus lookahead before issuing 16-byte wide stores; on an invalid code unit, scalar code resumes at that batch, preserving the exact first-error index and output prefix — the _with_errors contract is unchanged.

Performance

Recorded verification runs (x86-64), 32768-scalar inputs:

Testing

• Full test suite passes, with extended convert_utf32_to_utf8_with_errors coverage pinning the first-error index and output prefix across the new batch boundaries.
• Verified across implementations including the forced-fallback path; tests also pass on arm64 (which dispatches to unchanged kernels).

Full verification record: https://app.perfloop.ai/t/oss/case_nqbvgjwvc2