Add benchmarks for Parquet struct leaf-level projection pruning

[friendlymatthew]

Rationale for this change

This PR adds benchmarks that measure the perf of projecting individual fields from struct columns in Parquet files. #20925 introduced leaf-level projection masking so that select s['small_int'] on a struct with large string fields only reads the small integer leaf, skipping the expensive string decoding entirely

3 dataset shapes are coevered, each with ~262K rows of 8kb string payloads: a narrow struct (2 leaves), a wide struct (5 leaves), and a nested struct. Each shape benchmarks full-struct reads against single-field projections

[adriangb]

Oh how I wish we could write SQL benchmarks easily... adding to #21165