backport: Allow dynamic filter pushdown for left join

[discord9]

Summary

Backport upstream DataFusion's HashJoin dynamic filter support for join types whose probe side is safe for ON-clause filter pushdown.

This follows the original DataFusion feature behavior from apache#20447: HashJoinExec now gates self-generated dynamic filters with JoinType::on_lr_is_preserved() instead of a Greptime-specific Inner | Left rule.

The dynamic filter is still built from the left/build side join keys and applied to the right/probe side.

Note on null-aware LeftAnti

This backport intentionally keeps the behavior aligned with upstream DataFusion and does not add Greptime-specific handling for null-aware LeftAnti / NOT IN semantics.

In particular, this PR does not introduce an extra null_aware guard in HashJoinExec::allow_join_dynamic_filter_pushdown(). Any semantic caveats around null-aware anti joins are kept consistent with the original upstream feature and are considered out of scope for this backport.

What changed

• Add JoinType::on_lr_is_preserved() to datafusion-common.
• Reuse that helper from the optimizer's ON-clause filter-pushdown logic.
• Change HashJoinExec::allow_join_dynamic_filter_pushdown() to check whether the right/probe side is preserved for ON-clause pushdown.
• Update dynamic-filter sqllogictest expectations for the newly enabled join types.

Why this is safe

The self-generated HashJoin dynamic filter only filters the probe side by build-side join-key values.

For a physical LEFT JOIN, this is safe because rows on the right/probe side whose keys are not present on the left/build side cannot contribute to the output, while unmatched left rows are still preserved.

More generally, this backport keeps the upstream DataFusion rule: enable the feature when the probe side is preserved according to on_lr_is_preserved().

Validation

• cargo fmt --all
• cargo test -p datafusion-physical-plan test_allow_join_dynamic_filter_pushdown_gate -- --nocapture
• cargo test -p datafusion-physical-plan test_on_lr_is_preserved -- --nocapture
• cargo test -p datafusion-sqllogictest --test sqllogictests -- dynamic_filter_pushdown_config
• GreptimeDB integration check with this DataFusion patch:

cargo test -p tests-integration remote_dyn_filter_test -- --nocapture

The GreptimeDB integration test verifies:

• HashJoinExec: mode=CollectLeft, join_type=Left
• right-side MergeScanExec receives non-empty dyn_filters
• unmatched left row is preserved in LEFT JOIN output