[FEA] Expose filter_join_indices in Java

Is your feature request related to a problem? Please describe.

Java callers can reuse an equality hash table across probe tables with HashJoin, but cannot compose the resulting gather maps with an AST predicate. As a result, callers using Table.mixedInnerJoinGatherMaps or Table.mixedLeftJoinGatherMaps rebuild the hash table on every call.

libcudf already implements the desired composition: use cudf::hash_join::{inner,left}_join, then pass the resulting indices to cudf::filter_join_indices. The mixed inner/left/full implementation was rewritten this way in #23012, but filter_join_indices has no Java/JNI binding.

Describe the solution you'd like

Add a Java/JNI binding for cudf::filter_join_indices that accepts:

• left and right input GatherMaps;
• left and right tables containing the columns used by the predicate;
• a CompiledExpression;
• the join kind, supporting at least inner and left joins.

It should return newly allocated left and right GatherMaps without modifying or taking ownership of the input maps. The JNI implementation can borrow the gather-map addresses, wrap them in cudf::device_span<cudf::size_type const>, and call the existing libcudf API.

This would allow callers to compose the existing reusable-hash APIs with filtering:

try (HashJoin rightHash = new HashJoin(rightKeys, compareNullsEqual)) {
  GatherMap[] equalityMaps = leftKeys.innerJoinGatherMaps(rightHash);
  GatherMap[] result = Table.filterJoinGatherMaps(
      equalityMaps[0], equalityMaps[1],
      leftConditional, rightConditional,
      condition, JoinKind.INNER);
}

Tests should cover:

• reusing one HashJoin across multiple probe tables;
• equivalence with mixedInnerJoinGatherMaps and mixedLeftJoinGatherMaps;
• duplicate equality keys and a left row whose conditional matches all fail;
• nullable predicates;
• empty and mismatched gather maps.

Describe alternatives you've considered

Adding HashJoin overloads to the existing mixed*JoinGatherMaps methods would solve the immediate hash-table rebuild, but would preserve the monolithic mixed-join API and prevent the filter from being composed with other join implementations such as sort-merge join.

Spark RAPIDS JNI provides a similar custom primitive, JoinPrimitives.filterGatherMapsByAST, but cuDF Java should bind the existing libcudf primitive instead of duplicating its native implementation.

Additional context

This addresses hash-table reconstruction and allocator churn. It does not eliminate the materialized equality-join gather maps introduced by the join-then-filter composition; reducing that peak memory usage would be separate work.

The libcudf API is already present in 26.08 and current main. A Java/JNI binding in 26.10 would unblock callers upgrading from 26.06 that need reusable hash joins for mixed inner and left joins.

Related: #22124, #23012.