Is your feature request related to a problem or challenge? Please describe what you are trying to do.

Parquet writing performance is not very good. The arrow_writer microbenchmark shows throughput for a batch of primitives to be around 200MiB/s. For a column-oriented format that seems rather low, but a profiling run shows no obvious single bottleneck.

Describe the solution you'd like

Investigate and improve the performance.

• Optimize counting of values and nulls
• Avoid asserts called in a loop in BitWriter::put_value
• Avoids bounds checks in flush_bit_packed_run
• Optimize iteration in LevelInfoBuilder::write_leaf
• Avoid cloning null buffer or recalculating logical nulls
• Do not collect non_null_indices and gather these into a new Vec for non-nullable arrays
• Optimize writing bit-packed runs (bit width is 1 for levels most of the time, always writes 8 values except for last run)
• Change get_min_max to check logical/converted types outside of loop

Describe alternatives you've considered

Additional context