fix(core): close gaps for PyPaimon native Parquet writers

[JingsongLi]

Purpose

Close the remaining core and binding gaps exposed by enabling PyPaimon native Parquet writers for data-evolution tables, merge engines, nested input and primary-key value statistics.

Companion PyPaimon PR: apache/paimon#10202 (Native CI follows Rust main, so this Rust PR merges first).

Brief change log

• Delegate Python batch validation and normalization to TableWrite, allowing compatible nested child names and fixed-size binary input. Preserve schema-error recovery for the shared format-table writer path.
• Apply the core row-kind filter to both generated and explicit _VALUE_KIND input before bucketing and changelog writing. Accept global ignore-delete and Java fallback aliases for aggregation, and allow rowkind.field with supported merge engines.
• Reuse FirstRowMergeFunction for multi-row key groups during flush so every retract is checked. Preserve Java reducer behavior for singleton groups and the insert-only fast path.
• Honor parquet.write-page-index.enabled for both data and changelog files, retaining row-group statistics when page indexes are disabled and rejecting invalid settings before output creation.
• Parse bare VARCHAR, BINARY and VARBINARY names with Java's default length of 1, including nested Variant shredding schemas.

Tests

• Rust targeted tests: KV writer (27), aggregation options (21), row-kind schema validation (7), row-kind integration (21), table writer (69), types (22), Variant shredding (7), Parquet writer and physical data/changelog page-index regression.
• Python binding write tests: 29 passed, including normalization, explicit row kinds and schema-error recovery.
• cargo fmt --all --check.
• cargo clippy --locked --all-targets --workspace --features fulltext,vortex -- -D warnings.
• PyPaimon integration results will be recorded in the companion PR.

API and Format

Python write signatures remain unchanged; normalization and filtering live in Rust core. Adds a CoreOptions accessor for the existing Parquet option. Storage remains standard Paimon/Parquet.

Documentation

No new user-facing option: this implements existing Java/PyPaimon contracts.

[leaves12138]

Pushed 0684f18 to fix the two stale unit-test assertions without reverting the intended implementation behavior:

• The aggregation schema rejection test now checks genuinely unsupported options rather than the newly supported global ignore-delete option.
• The Format Table test now expects an empty successful prepare after a rejected schema. Added a regression proving that valid rows accepted before and after an invalid batch are preserved and committed.

Validation: cargo test --locked -p paimon --lib passed (3398 passed, 6 ignored); the RowKind and physical Parquet page-index integration tests passed (22 tests); cargo fmt --all --check and git diff --check passed. The new GitHub CI run is still in progress.

[QuakeWang]

+1

[leaves12138]

LGTM after re-review of 0684f18. The two outdated assertions are corrected, and the additional Format Table regression verifies that a schema error preserves previously accepted rows and allows subsequent valid writes. No further blocking issues found in the reviewed RowKind filtering, first-row reducer, Parquet page-index, type-default, or Python binding changes. Verified all 14 checks are successful; the Ubuntu core suite reports 3496 passed and 6 ignored, and the Python binding suite reports 283 passed, including all 29 write tests. The previously failing tests and the new recovery test all pass in CI.