bench: add the shuffle benchmark cases that the #5905 review found missing #7080

What is the problem the feature request solves?

The shuffle benches in native/shuffle/benches do not cover several shapes that the performance review in #5905 flagged, so changes to those paths cannot show a win or catch a regression. What exists today: shuffle_writer.rs (codecs None/LZ4/Snappy/Zstd, a high-partition group, partitioning placement, schema encoding, IPC context), shuffle_reader.rs (codecs none and lz4, dictionary blocks, schema parse) and row_columnar.rs (primitive, struct, list and map conversion).

Describe the potential solution

Add the missing cases, each as its own small PR:

• Writer: end-to-end wide schema (100+ columns); List, Struct and Map through interleave; a spill run under a memory limit; a compressed high-partition case (the existing high-partition group uses codec None).
• Reader: zstd and snappy blocks; List and Map blocks; a cached-schema variant against decode_block_uncached for nested schemas.
• Row to Arrow (row_columnar.rs): strings (plain and dictionary-encoded), List, Map<Utf8, _> and nullable data. The existing list case list_conversion/elements_100 has data-dependent output sizes that [EPIC] Nested-type conversion performance: C2R, R2C and JVM shuffle slower than Spark #6855 asks someone to look at first.

Record results in the PR description with cargo bench -- --save-baseline.

Additional context

"Benchmark coverage to add" section of #5905. Good first issue; the individual items are independent. Related: #6856, #6855.