GUASI update

GUASI update

The GUASI library has been changed to avoid using pthread defaults stacks
(that burns 8MB of VM area for each thread). Also, a new paramter has been
added to guasi_create(), to allow the caller to specify a stack size.
I noticed a different behaviour of FIO in randrw mode, from GUASI to libaio.
In GUASI, the amount of data read is about the same +-2-3% of the amount
of data written. Whereas in libaio, there can be a huge difference between
the two (up to 2-3 times the other). This is consistent.
Besides that, GUASI performance follows pretty much the libaio one, up to
an io-depth of 1000-1500, after that shows about 10-15% less performance.

Signed-off-by: Jens Axboe <jens.axboe@oracle.com>