[core] Align delete and update nullability with Java

[JingsongLi]

Purpose

Align existing Python Native write/update behavior with current Java, preserving logical NOT NULL constraints while allowing valid deduplicate DELETE records and unused update inputs.

Companion Python PR: apache/paimon#10666.

Changes

• Derive a nullable physical KV value schema for deduplicate tables without cross-partition updates, matching PrimaryKeyFileStoreTable.store(). Preserve primary-key/sequence fields and the logical table schema.
• Preserve physical root nullability in the Paimon fields supplied to the Parquet writer, including data and INPUT changelog files. Preserve field IDs, descriptions, defaults and nested nullability.
• Resolve row kinds and enforce logical top-level NOT NULL checks before the optional row-kind filter, including when no ignore option is enabled. Other merge engines and cross-partition writes retain strict checks.
• Reuse the existing write-input Arrow importer for update/upsert/MERGE sources. Preserve actual NULLs, sliced buffers and metadata until Core selects target payloads; selected target constraints remain enforced in Core.
• Keep one common transport schema across the chunks of an Arrow Table by unioning relaxed root nullable flags without copying payload buffers. Core MERGE retains its strict schema check; heterogeneous iterable upsert inputs retain their supplied field sets.
• Update the existing invalid-row-kind test to assert rejection at write time and no staged output.

Verification

• Review follow-up: real PyArrow-to-Core MERGE reproduces the old mixed-chunk schema failure, then passes with NULL chunks in either order, NULL ON keys, unmatched NULL payloads, and unused dictionary values that decode to NULL. Source schema/field metadata remains intact. Rust 1.98.0 binding all-target Clippy, rustfmt and diff checks pass.

• New retained-delete integration suite: 3 passed. Covers explicit/configured row kinds, STRING/LargeString, automatic/precomputed buckets, separate INSERT and DELETE commits, sequence ordering, physical Parquet root fields, INPUT changelogs, value predicates that must not resurrect deleted rows, managed BLOB files, and required-field/other-engine/cross-partition negatives.

• Existing filtered-write integration suite: 4 passed.

• Python-binding Rust library: 104 passed. New source-transport and real upsert tests first fail on the old importer, then pass, including discarded duplicate values and invalid selected target NULLs.

• Rust core library: 4319 passed, 7 ignored. Pinned Rust 1.98.0 core/binding all-target Clippy with -D warnings, rustfmt and diff checks passed.

• Current Java reference tests TableWriteTest.testDeduplicateDeleteWithPrimaryKeyOnly and testDeleteRequiresSequenceField: 2 passed.

• Companion Python regressions cover ordinary/dedicated-BLOB target constraints, physical schema flags, and real commit/read results. Local Python tests did not rebuild or install the Native extension.

No new dependency, public API, version probe or commit-failure cleanup is introduced.

[leaves12138]

LGTM