Add flat vs. struct field projection benchmarks

[friendlymatthew]

Rationale for this change

This PR adds a benchmark comparing top-level column access against struct field access for the same logical data

#20925 introduced leaf level projection masking so that projecting a single struct field skips decoding its siblings. #21180 added benchmarks measuring that improvement across different strcut shapes. But neither benchmark answers how struct field access compare to reading the same column at the top level. Without that baseline, it's hard to know how much overhead the struct access path itself adds