feat: transitive predicate propagation across multi-table join chains

[xanderbailey]

Which issue does this PR close?

• Closes #.

Rationale for this change

Right now, PushDownFilter does single-hop predicate inference through join equi-conditions. If you have a.x = b.y as a join key and WHERE a.x > 5, it'll infer b.y > 5 and push it down. Great.

But for multi-table join chains like:

SELECT * FROM a
  JOIN b ON a.x = b.y
  JOIN c ON b.y = c.z
WHERE a.x > 5

...it never derives c.z > 5. The problem is that b.y > 5 gets pushed below the first join and becomes invisible when processing the second join. This means table c gets scanned without any filter, which is a missed optimization that matters for any query joining 3+ tables on shared keys.

Spark, and Calcite all do transitive predicate closure -- DataFusion should too.

• Spark: InferFiltersFromConstraints in the Catalyst optimizer
• Calcite: JoinPushTransitivePredicatesRule

What changes are included in this PR?

Rather than adding a new optimizer rule, this extends the existing PushDownFilter rule with equivalence-class-based inference.

The key additions in push_down_filter.rs:

• ColumnEquivalences -- a simple union-find over Column that tracks which columns are transitively equal
• collect_descendant_equalities() -- walks the plan subtree collecting column equalities from descendant INNER join ON clauses. Only collects from INNER joins (outer join equalities don't unconditionally hold). Stops at projections, aggregates, limits, and other nodes that change column identity or row cardinality.
• infer_predicates_from_equivalence_classes() -- for each single-column predicate, generates equivalent predicates for all columns in the same equivalence class
• The existing infer_join_predicates() is modified to build equivalence classes from the current join's keys + descendant joins + WHERE-clause column equalities, then use them for inference

The existing ON-filter inference path (infer_join_predicates_from_on_filters) is kept as-is since it correctly handles join-type-specific directionality for ON clause predicates.

Also adds push_down_filter_transitive.slt with end-to-end SQL tests that verify both EXPLAIN plans and query correctness.

Are these changes tested?

Yes -- both unit tests and sqllogictests:

Unit tests (new tests in push_down_filter.rs):

• 3-table and 4-table INNER join chains
• Mixed INNER + LEFT join chain
• WHERE-clause column equality (WHERE a.x = b.y AND a.x > 5)
• FULL JOIN (verifies correct handling at the current join level)
• Self-join with aliases
• Projection boundary (verifies recursion stops correctly)
• Complex ON expressions (verifies a.x + 1 = b.y is not treated as column equality)
• ColumnEquivalences unit tests (basic operations + path compression)

SQL logic tests (push_down_filter_transitive.slt):

• 3-table chain: EXPLAIN shows filter pushed to all tables + correct results
• 4-table chain: same
• Mixed INNER + LEFT JOIN: filter propagates through both
• Correctness checks with/without filters

Are there any user-facing changes?

No API changes.