Thish fixes Suparna's fsx and aio-stress and Dave Olien's concurrent
mke2fs problems.

It moves the hash adding into as_add_request.  This fixes a small bug
where a request could be added to the merge hash after being put on the
dispatch list. 

It also moves responsibility for handling rbtree aliases (requests with
the same start sector) out of as_add_arq_rb and into its caller.  This
cleans up a layering problem which was causing the state machine to go
silly.  Now instead of forcing the alias out onto the dispatch list, we
chain the new request behind it, and they get sorted out at dispatch
time.