On Linux, device IO goes through the buffer cache by default. This makes
frisbee run really fast...until it closes the output device. Then it sits
for minutes while it flushes disk data out of the cache. This is
technically okay, but wasteful, since frisbee allocates its own memory
for caching disk write data. By using direct IO on the output device,
writes do not go through the cache.

Aren't two caches better than one? No. They can compete for memory and it
just causes an extra data copy. Frisbee is faster when using O_DIRECT.

Note that this change is complicated somewhat because Linux requires that
the IO buffer for an O_DIRECT opened file be sector aligned. So we play
some games to do this.

This should have no effect on FreeBSD where device writes don't go through
the buffer cache.