This is Al's implementation of the Orlov block allocator for ext2.

At least doubles the throughput for the traverse-a-kernel-tree
test and is well tested.

I still need to do the ext3 version.

No effort has been put into tuning it at this time, so more gains
are probably possible.