perf(table): push row-range pruning into manifest scan

[XiaoHongbo-Hope]

Summary

Push row-range pruning into the data-manifest scan for data-evolution tables. Fast and Full global-index modes now determine selected row ranges before reading data manifests, so unrelated manifest files and entries can be skipped before merge, grouping, and statistics pruning.

Changes

• Evaluate Fast and Full global-index lookups before data-manifest I/O; keep Detail mode's post-manifest fallback for unindexed ranges.
• Prune whole manifests by row-id bounds, then prune decoded entries before merge and data-evolution grouping.
• Apply the same row-range path to snapshot, delta, and changelog planning while preserving deletion-vector handling.
• Fail open when row-id metadata is missing or invalid, and retain downstream pruning as a correctness safeguard.
• Add scan-trace counters for manifest files and entries pruned by row ranges.
• Cover manifest-level pruning, entry-level pruning, Detail fallback, invalid metadata, and rolled BLOB/vector scan-to-read behavior.

Testing

• cargo fmt --all -- --check
• cargo clippy --locked --release -p paimon --lib --tests --no-deps -- -D warnings
• cargo test --locked --release -p paimon --lib (1842 passed, 1 ignored)
• Same-snapshot A/B on a real data-evolution table: median planning time improved from 2.203 s to 1.342 s with identical planned files and read results.

Notes

Selected-range reads for rolled dedicated BLOB/vector files rely on the reader support added by #592.

[JingsongLi]

Any filtering that can be done in advance should be done so, and should even be moved down to the manifest reading phase.

[QuakeWang]

LGTM

[JingsongLi]

Just for deletion-vectors? If it is. It needs to be in DV mode to be read. And it should be filtered so that only DV remains.

[XiaoHongbo-Hope]

Just for deletion-vectors? If it is. It needs to be in DV mode to be read. And it should be filtered so that only DV remains.

Updated.

[JingsongLi]

+1

[leaves12138]

Found one correctness edge case in the empty row-range path.

[leaves12138]

Could we short-circuit an empty manifest_row_ranges here? When the global index returns Some(vec![]), a manifest/file with missing or invalid row-id metadata makes both row-range pruning helpers fail open. The file then reaches split_row_ranges_for_files, which returns DataInvalid because row_id_range() is None. As a result, a valid zero-hit Fast-mode lookup on a legacy data-evolution table fails instead of returning an empty plan. I reproduced this by planning with Some(Vec::new()) against a committed file whose first_row_id is None; plan() returns Cannot apply selected row ranges.... Please return an empty plan before reading/planning data manifests when the selected range set is empty, and add a regression test.