Add benchmarks for reading struct arrays from parquet

[jhorstmann]

Which issue does this PR close?

• Closes [Parquet] Add benchmarks for reading struct arrays from parquet #9209.

Rationale for this change

Improve the performance benchmark coverage.

What changes are included in this PR?

Are these changes tested?

Benchmark results on my machine, without target specific optimizations:

arrow_array_reader/Int32Array/plain encoded, mandatory, no NULLs
                        time:   [8.5123 µs 8.5218 µs 8.5321 µs]

arrow_array_reader/Int32Array/plain encoded, optional, no NULLs
                        time:   [12.843 µs 12.863 µs 12.885 µs]

arrow_array_reader/Int32Array/plain encoded, optional, half NULLs
                        time:   [64.977 µs 65.108 µs 65.267 µs]

arrow_array_reader/struct/Int32Array/plain encoded, mandatory struct, optional data, no NULLs
                        time:   [14.456 µs 14.473 µs 14.491 µs]

arrow_array_reader/struct/Int32Array/plain encoded, optional struct, optional data, no NULLs
                        time:   [173.14 µs 173.68 µs 174.23 µs]

arrow_array_reader/struct/Int32Array/plain encoded, mandatory struct, optional data, half NULLs
                        time:   [65.636 µs 65.780 µs 65.906 µs]

arrow_array_reader/struct/Int32Array/plain encoded, optional struct, optional data, half NULLs
                        time:   [433.95 µs 434.29 µs 434.67 µs]

I think these show that a mandatory struct has negligible overhead compared to a non-nested primitive column. But an optional struct, even if its definition level is always valid, has a significant overhead.

Are there any user-facing changes?

no

[alamb]

Makes sense to me -- thank you @jhorstmann

I also verified that the benchmark shows non trivial time being spent in definition levels

I can't wait to see what else is coming