test: add comprehensive JVM shuffle benchmarks

Add jvm_shuffle.rs benchmark that covers the full range of data types
processed by `process_sorted_row_partition()` in JVM shuffle:

- Primitive columns (100 Int64 columns)
- Struct (flat with 5/10/20 fields)
- Nested struct (2 levels deep)
- Deeply nested struct (3 levels deep)
- List<Int64>
- Map<Int64, Int64>

This replaces the old row_columnar.rs which only tested primitive columns.

These benchmarks help measure the performance of the row-to-columnar
conversion used by CometColumnarShuffle when writing shuffle data.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Author: andygrove

native/core/Cargo.toml

@@ -122,7 +122,7 @@ name = "bit_util"
 harness = false
 
 [[bench]]
-name = "row_columnar"
+name = "jvm_shuffle"
 harness = false
 
 [[bench]]