[python] Use native upserts for equivalent Arrow input layouts

[JingsongLi]

Purpose

Use Native execution for Arrow upsert inputs that PyPaimon already supports: large-string match keys, partition fields and nested values, plus equivalent fixed-size binary keys. The current exact type comparison sends these operations back to Python even though they represent the same logical table values.

Depends on apache/paimon-rust#1144. Native CI continues building apache/paimon-rust main, so the Rust change must merge first.

• Reuse the existing Arrow schema compatibility policy for supplied upsert columns. Core owns normalization, partition-scoped matching, last-write-wins deduplication and file writing. Numeric conversions, different timestamp zones and other non-equivalent types keep their existing dispatch behavior.
• Reuse read_arrow_schema when combining DE column bunches. File readers already return runtime-nullable nested types; reapplying the stricter logical schema could make an otherwise valid read fail.
• Reuse the core-backed write normalizer at the Python row-ID writer boundary to restore declared nested write types and reject visible NULLs in required children. Keep NULL parent ROW values valid.
• Close both the mapped batch iterator and its underlying merged reader on validation or writing errors, even while the exception traceback is retained.
• Add Native/Python parity tests covering batch/stream upserts, multiple Arrow chunks, nested ROW/ARRAY/MAP strings, duplicate and NULL keys, partition isolation, partial writes, row identity, REST-backed writes, and required nested fields in row-ID and predicate updates.

The scope is read/write/update execution. There are no catalog metadata transport changes or new public APIs, options or dependencies.

Tests

Built the Rust bindings from apache/paimon-rust#1144 and ran with Native planning, reading, writing and committing enabled:

• New layout cases plus chunked row-ID writer tests: 42 passed before the final reader-close regression was added.
• Final regression selection (native_upsert_layout_test, table_update_by_row_id_chunked_test, table_update_test, data_evolution_test, native_blob_update_test): 383 passed, 24 subtests passed.
• Broader upsert/nested update/BLOB/DE/concatenated reader selection: 568 passed, 2 skipped, 24 subtests passed.
• Flake8 for every changed Python file and git diff --check.
• The original Native fallback and core LargeUtf8 rejection were reproduced before the fixes. The new reader-close test failed in both validation-error and writer-error cases before explicitly closing the inner iterator.
• Independent review of type eligibility, nested constraints and resource ownership; the identified reader-close issue was repaired and re-reviewed.

[leaves12138]

LGTM