This patch fixes the deliberate BUG_ON() on failed sgtable allocations.
I left that in because I was too lazy to fix it at the time...

Basically make scsi_init_io() return 0 on failure (like before) but this
time allow us to recover. scsi_init_io() will end_request on a fatal
condition. So now just mark the device/host as starved if needed, and
leave the request at the front of the queue for next service.