This does 4 things.  Fixes the oversight James found about cmd_per_lun.
Makes scsi_adjust_queue_depth() implement the usage of tagged that I
described in my email to linux-scsi.  Updates some comments to highlight
things that are going to go away shortly.  Removes the call to
host->select_queue_depths() entirely since drivers that implement this
will try to set SDptr->queue_depth to the desired depth, which in reality
will do nothing but confuse the mid layer since that is now the actual
allocated command counter, so it won't make the mid layer allocate more
commands, it will make the mid layer think more commands have been
allocated.