Call blockdev --rereadpt after sfdisk commands.

On fast machines sfdisk has some sort of race where it
fails to re-read the partition table it has just created
(it's not clear if this is a race in sfdisk, the kernel or
some other component).

This commit works around the problem by calling
blockdev --rereadpt after sfdisk operations, which
experience shows is enough to stop the problem from
happening.