Use RoaringTreemap for MergeRows cardinality tracking #6608

Problem

Native MergeRowsExec uses HashSet<i64> for cardinality validation, while Spark uses Roaring64Bitmap.

As measured in the #5318 review, for 5M row IDs:

Since this state cannot spill, native MERGE can hit ResourcesExhausted where Spark still fits in memory.

Proposed solution

Replace HashSet<i64> with RoaringTreemap while preserving duplicate detection, existing cardinality semantics, and DataFusion memory accounting.

Add coverage for dense/shuffled IDs, cross-batch duplicates, and bounded-memory cases.

Related: #5318, #5122