[python] Plan chunk-shuffled native reads with IndexedSplit

[JingsongLi]

What changed

• pass chunk-shuffle seed, live-row chunk size, and scan-level worker shard directly to the Rust table scan
• deserialize native chunk plans through the standard Java-compatible IndexedSplit wire format
• use IndexedSplit.row_ranges consistently in Python planning and reading
  • row-tracked and data-evolution tables: stable global row IDs
  • tables without row tracking: split-local physical positions over file order
• expand deletion vectors into live-only ranges, making row_count and merged_row_count the exact sum of selected ranges
• select the coordinate system from table options and map ranges to each file in RawFileSplitRead, including DV position remapping and multi-file chunk boundaries
• remove the temporary native metadata facade and exact-count overrides from IndexedSplit / SlicedSplit
• fall back to the complete Python planner when the installed binding does not expose native chunk planning

Coverage

The matrix covers append, row-tracked append, and data-evolution tables; all four Python/Rust planner-reader combinations; deletion vectors including fragmented alternating positions; multi-file chunk boundaries; null and non-null partitions; signed and large seeds; worker sharding; projection; partition filters; time travel; empty plans; incremental reads; and stable split serialization.

Dependency

Depends on apache/paimon-rust#890.

Verification

• chunk-shuffle planner, reader mapping, and native chunk integration: 85 passed
• all native-plan and native-reader tests: 360 passed, 128 subtests passed
• split serialization, distribution, and temporal suites: 105 passed, 77 subtests passed
• Rust binding read suite: 87 passed
• PyTorch suite was not run locally because PyTorch is not installed

[leaves12138]

[P2] Remove the now-unused SlicedSplit import to unblock Python CI

This PR removes the remaining SlicedSplit assertions from paimon-python/pypaimon/tests/native_plan_distribution_test.py, but leaves the import at line 35. The required lint step now fails with:

pypaimon/tests/native_plan_distribution_test.py:35:1: F401 'pypaimon.read.sliced_split.SlicedSplit' imported but unused

Reproduced locally with python -m flake8 --config=dev/cfg.ini pypaimon/tests/native_plan_distribution_test.py from paimon-python. The Python 3.10 CI job reports the same error: https://github.com/apache/paimon/actions/runs/35521350138/job/106105908581 . Please remove the unused import and rerun CI.

I am continuing paired-runtime correctness validation. Compatibility with intermediate, unreleased native split implementations is not being treated as a separate blocker.

[leaves12138]

LGTM for the Python-side changes through fb79b66. The latest commit fixes the unused-import CI failure reported in my earlier review. I found no remaining blocking issue specific to this Python patch.

Validation with the actual companion #890 runtime (724d2d685b852519926e4fff181526c1efe7a421):

• 543 targeted PyPaimon tests passed, covering native planning, chunk shuffling, DVs, split serialization, reading, distribution, and temporal reads.
• 88 Rust Python-binding read tests passed.
• After the final import-only commit, the 14 native-plan distribution tests and flake8 on all changed Python files passed.

This approval is scoped to this Python PR, not the Rust dependency: I separately reported a reproducible ordinary row-ID range-read regression on apache/paimon-rust#890, which should be fixed before the paired feature is released. Compatibility with intermediate, unpublished native split implementations is not a blocker for this review.

[leaves12138]

LGTM at 154af1a, validated together with apache/paimon-rust#890 at 7596b12d00b531983ab932a7a2412f8ceca65e46.

The latest fix addresses the remaining mixed-runtime row-loss issue: row-tracked append chunk ranges are generated as global row IDs, and RawFileSplitRead maps them back to file-local physical positions before reading. Non-row-tracked tables retain split-local coordinates. The added cross-runtime regression test exercises both mixed directions with separate planner/reader configurations.

Local validation:

• The original eight-case planner/reader matrix passes with complete row coverage.
• 546 targeted PyPaimon tests and 89 Rust Python-binding read tests pass.
• 18 additional mixed-runtime cases pass across append/data evolution, row tracking, deletion vectors, chunk sizes 1/3/5, worker assignment, partition-filter projection, and snapshot reads.
• flake8 passes on all changed Python files.

No remaining blocking issue found. The GitHub Python CI matrix was still running when checked; this approval does not imply that remote CI has completed.