Add a benchmark for large file scans versus smaller files

[jonkeane]

When scanning and doing operations on files, it would be nice to know if it's more efficient to have one large parquet (for example) file per partition, or to have more smaller files.

[lidavidm]

This would be good to have. For Parquet specifically, looking at row group sizes may also be interesting - we can potentially get more parallelism with smaller row groups, but if you're reading only a few sparse columns of many, and you're on something like S3, small row groups also mean you have to make lots of small reads which is not an ideal I/O pattern.