[python][ray] Add periodic commits to write_paimon

[XiaoHongbo-Hope]

Purpose

write_paimon normally publishes one snapshot after the entire Ray Dataset finishes. For a job that runs for hours or days, a late failure leaves none of that write visible, so users must manually split the input and call write_paimon batch by batch.

This PR makes that batching internal. With commit_mode="incremental", completed primary-key groups are committed at a target time interval. If a later group fails, earlier commits remain visible.

This is periodic commit, not source checkpoint recovery. Retrying reruns the Ray Dataset from the beginning and upserts already committed keys again.

Changes

• add commit_mode="incremental", commit_interval_seconds, and update_cols to write_paimon;
• periodically commit completed fixed-bucket primary-key groups;
• update existing keys and insert new keys using merge-engine=partial-update;
• retain completed commits after a later worker failure;
• allow concurrent partial updates to different fields.

The target must be a fixed-bucket primary-key table using merge-engine=partial-update. Incremental mode requires Ray 2.33 or later. Concurrent jobs updating the same field must not rely on a deterministic winner.

Tests

• real Ray E2E for periodic commits, upserts, worker failure, full retry, concurrent partial updates, 4096 buckets, partitioned composite keys, and compaction;
• deterministic regression for a materialized Dataset whose group writes exceed the interval;
• Hangzhou staging DLF + OSS E2E covering two periodic snapshots, concurrent updates to different fields, injected failure, full retry, and preserved non-updated columns;
• uncertain-commit callback, Ray sink, shuffle, and table commit regressions;
• project flake8 and git diff --check.

Latest local validation: 100 passed with Ray 2.56.1.