[python][ray] Avoid routing shuffle in self-merge updates

[XiaoHongbo-Hope]

Purpose

Ray _ROW_ID self-merge already avoids the source-target join, but its update path still sends every transformed row through a _FIRST_ROW_ID hash shuffle before writing. This adds avoidable network, object-store, and scheduling overhead for large column updates.

Changes

• pin the target scan plan used by self-merge updates
• unpack planned splits into complete first_row_id file groups and dispatch
  each group directly to a bounded Ray task
• read, apply matched clauses, and stage row-id updates within that task
• collect normal CommitMessages and commit once on the driver
• stop scheduling new work after a task failure and abort files staged by completed tasks

This bounds task materialization to one logical file group rather than a packed split. The public API is unchanged. Callable assignments from #9327 reuse the same path; general source-target merge and self-merge delete paths are unchanged.

Tests

• self-merge updates succeed when Dataset.groupby() is disabled
• a packed split is separated into complete logical file groups
• callable read_columns and VARIANT transformations work without the shuffle
• a later group failure aborts files staged by earlier groups
• ray_data_evolution_merge_into_test.py: 115 passed
• ray_update_by_row_id_test.py: 16 passed
• flake8, compileall, and git diff --check passed

Local filesystem microbenchmark shaped like the reported workload (Ray 2.56.1, 4 CPUs, append-only data-evolution table with row tracking, 100,000 rows, 20 files, 65 clips per file, and 1–3 KiB VARIANT payloads): master with the routing shuffle took 5.43/5.57 s; this PR took 1.45/1.50 s (about 3.7x). The callable performs variant_get, negates four DOUBLE paths, and writes them with variant_replace. All 60,060 matched rows and 39,940 unmatched rows were verified after every run. This is a local directional result, not a production-table claim.