Add direct intern() benchmark for vectorized vs row-based GROUP BY

Adds a fair apples-to-apples benchmark that directly calls
GroupValues::intern() with identical Int32 data for both
GroupValuesColumn (vectorized) and GroupValuesRows (row-based).

This eliminates the previous confounding factors (different data types,
SQL/planning overhead) and confirms the regression reported in #17850:
row-based is 16-19% faster at low cardinality (64 groups), with a
crossover at ~200K-500K groups where vectorized becomes faster.

Experiments:
- Issue #17850 reproduction (3 cols, 64 groups, 1M-50M rows)
- Low cardinality sweep (8-4096 groups)
- Batch size sensitivity (1K-32K)
- Column count scaling (2-10 cols)
- High cardinality scaling (1M groups)
- Group count sweep (16 to 1M groups)
- Random vs sequential data patterns

Author: Nathan Bezualem

datafusion/physical-plan/Cargo.toml

@@ -116,3 +116,7 @@ required-features = ["test_utils"]
 [[bench]]
 harness = false
 name = "dictionary_group_values"
+
+[[bench]]
+harness = false
+name = "multi_group_by"