Once upon a time, there was a bdrv_pwrite that actually wrote single bytes to
the file... However, today it is emulated by a read-modify-write cycle which
aligns the request to sector size. This is slow. And we don't need it: qcow2
often has the complete sector in memory, we don't need to read it from the disk
again.
These patches change the writes to L1 tables, L2 tables and refcount blocks to
write complete sectors instead of single entries.
This series depends on the qcow2 split to apply cleanly.