[core] Support logical Arrow keys and selected upsert values

[JingsongLi]

Purpose

Expand Native row-level updates for Arrow input shapes already supported by PyPaimon. This covers action-specific upsert payload conversion, lossless logical key matching, and Dictionary keys in MERGE. Matching and conversion stay in Rust core; the Python bridge only qualifies supported inputs and transports batches.

Brief change log

• Select surviving upsert rows before converting payloads. Reuse row-ID update conversions for matched rows, ignore unwritten columns, and retain ordinary write validation for appended rows.
• Normalize upsert keys independently of their original payload arrays. Support exact integer widths, decimal scales, date/time/timestamp units, NULL-typed arrays, and Dictionary encodings. Reject overflow, lost precision, and timezone-awareness changes.
• Preserve original key arrays for action-specific write validation: a key conversion accepted for matching must not implicitly relax append validation.
• Decode MERGE keys before row encoding and SQL NULL filtering. Different dictionaries in separate batches cannot change key identity. Retain cardinality validation before conditions and the unconditional DELETE exception.
• Reuse existing row-ID conversions for matched Binary/LargeBinary values, including nested MAP/ROW and constructor-shaped inputs. No upsert-specific conversion policy or writer flag.
• Reuse existing nested field projection and core conversion helpers. Upsert, MERGE and Blob map key layout normalization share the internal normalize_match_key helper; logical type and NULL matching rules remain operation-specific. No new public API or dependencies.

Tests

• cargo +1.98.0 test --locked -p paimon --lib key_normalization: 5 passed.
• cargo +1.98.0 test --locked -p paimon --lib upsert: 10 passed.
• cargo +1.98.0 test --locked -p paimon --lib table_merge_into_tests: 17 passed.
• cargo +1.98.0 test --locked -p paimon --lib table::update_input: 29 passed.
• cargo +1.98.0 test --locked -p paimon --lib table::data_evolution_writer: 19 passed.
• cargo +1.98.0 test --locked -p paimon --lib table::write_batch_normalize: 24 passed.
• cargo +1.98.0 clippy --locked -p paimon --lib --tests -- -D warnings.
• cargo +1.98.0 fmt --all -- --check and git diff --check.
• Built the Native extension from this branch and ran the companion PyPaimon regressions with Native plan/read/write/commit enabled: 1322 passed, 2 skipped. Exercised 1887 Native plans, 2236 reads, 809 writes, and 257 commits.

Coverage includes batch/stream upserts, last-write-wins across chunks, fan-out, partition isolation, NULL indices and NULL dictionary values, discarded invalid payloads, strict append validation, precision/overflow failures, Binary/LargeBinary conversion parity, Dictionary MERGE update/insert/delete/cardinality, and Pandas categorical input.

The companion apache/paimon#10629 also enables existing core Parquet vector/BLOB update support across row-ID, incremental, predicate, upsert, MERGE, and shard operations, with Python/Native parity coverage. Its Native CI requires this PR on Rust main.

API and Format

No public API or storage format change. Existing core row-ID conversion rules and Java-compatible write/file metadata semantics are reused.

Documentation

Clarified TableUpdate::upsert_by_arrow_with_key matched conversion and append-validation contracts and documented logical key normalization internally.

Follow-up review fixes

• Remove the upsert-specific Binary/LargeBinary restriction and its recursive policy plumbing. Matched values now directly use the established shared row-ID conversion implementation; ordinary append validation keeps its existing rules. Positive Python/Native batch and stream tests cover arbitrary bytes, NULLs, nested MAP/ROW values and constructor inputs, with unchanged row identities.
• Preserve Blob map transport and Java key serialization independently of value conversion: Dictionary keys and Dictionary payloads, exact integer widths, DECIMAL HALF_UP and TIME millisecond normalization. Normalization-induced duplicate keys still fail, and append types remain strict.
• Handle exact decimal zero/NULL keys across extreme scale differences without Arrow power-table errors or i8 scale arithmetic overflow. Nonzero overflow and precision loss remain errors.
• Add regression cases for those failures, NULL containers, mixed Blob map transport and unchanged row identities. Several independent review passes found no remaining confirmed issue after repairs.

[leaves12138]

LGTM