Add how much block I/O is requested (by load_block) to the Metrics struct.
Per block type (data/index/filter).
That allows us to benchmark read I/O without being dependent on the OS's reporting of disk I/O (so we can measure block I/O that comes from the page cache as well).