[java][python] Read Python-planned splits with the native reader

[JingsongLi]

Purpose

Keep PyPaimon's Rust physical reader available when the scan must be planned in Python. This extends native coverage to Python-planned DataSplit and IndexedSplit inputs, including Data Evolution row ranges and query-authorization wrappers.

Uses the cross-language split deserializer merged in apache/paimon-rust#889. pypaimon-rust has not been released, so this intentionally targets that API without a legacy compatibility shim. Vortex native read remains out of scope.

Changes

• implement a Java-compatible SplitSerializer v1 writer for Python DataSplit and IndexedSplit
• serialize the current DataSplit v9 / DataFileMeta v21 layout, preserving stats, deletion vectors, row ranges, external paths, streaming/raw flags, and optional vector scores
• bridge Python-planned splits through pypaimon_rust.datafusion.Split.deserialize before native reads; scores remain on the Python IndexedSplit while Rust receives the physical row ranges
• attach the planned snapshot id to Python DataSplits so stable cross-language serialization is self-contained
• align the Java and Python APIs on writeColsSequences / write_cols_sequences, while retaining _WRITE_COLS_SEQUENCES as the serialized field name
• support native reads for homogeneous query-authorization splits with Python-equivalent ordering: authorization filter, shared LIMIT, masking, then public projection
• widen native projection for authorization dependencies without exposing those columns to callers
• preserve safe Python fallback for unsupported split wrappers, mixed authorization contexts, nested or MAP-key authorization, and authorization filters over BLOB payloads
• preserve initial-stream visibility of uncompacted level-0 files by using the Python planner when Rust's full-snapshot batch semantics would hide them; incremental native planning remains enabled
• harden split decoding against truncated, negative-count, malformed-array, inverted-range, oversized modified-UTF, and trailing-byte inputs
• add real Python-plan-to-Rust-read integration coverage for ordinary, partitioned, Indexed/Data Evolution, authorization, streaming-batch, projection, parallelism, and fallback paths

Tests

• native read, split serializer, and query authorization: 114 passed
• real paired Rust-wheel native integration suite: 35 passed, including Python-planned and Indexed/Data Evolution reads
• Rust-plan level filtering suite: 15 passed, with 24 native plans exercised
• streaming and native incremental planning suites: 66 passed
• native Python coverage group: 266 passed, plus 128 parameterized subtests
• resolved-schema coverage group: 82 passed
• non-Vortex Data Evolution group: 185 passed, 1 skipped, 10 deselected, plus 29 parameterized subtests
• changed manifest round-trip test passed
• Java DataFileMeta / Data Evolution / global-index tests: 57 passed
• Flink serializer tests: 7 passed
• Spark copy tests: 3 passed
• python -m compileall -q pypaimon
• project flake8 configuration passed on every changed Python file
• paired Rust binding tests, clippy, and formatting are recorded in [python] Deserialize cross-language splits for native reads paimon-rust#889

[leaves12138]

Reviewed through 5877a49. The latest commit forwards snapshot_id from SlicedSplit and adds regression coverage, resolving the test failure found during review. I found no remaining blocking issues in the reviewed changes.

Checked the Python-to-native split conversion and serialization, query-authorization fallback, initial streaming behavior, and metadata compatibility.

Validation:

• 556 targeted Python tests passed with the real native runtime from apache/paimon-rust#889 (1cda72a4beacb94a835b45daedbfb338ac05fa86), including native chunk-shuffle and split-generator tests.
• 195 targeted Java serializer, compatibility, manifest, and Data Evolution tests passed; the subsequent commits only change Python code/tests.
• Changed Python files passed flake8, including the latest wrapper and test changes.

Remote CI is still running; this approval does not imply that all CI jobs have completed.