Move enqueueing of commands to bus_dmamap_load_ccb callback

Fix fundamental difference between FreeBSD and Linux. On Linux, your dma
load callback always happends before it returns, so drivers are written
to load the map, then submit to hardware. On FreeBSD, the callback may
be deferred and return EINPROGRESS. This means the callback is
responsible for queueing the request to the hardware is done after the
SGL list is created. Make a number of interrelated cahnages:

At the end of mpi3mr_prepare_sgls, add a call to mpi3mr_enqueue_request.

Split the hardware submission out from the end of mpi3mr_action_scsiio
and move it into a new routine mpi3mr_enqueue_request.

Move all error completion from the end of mpi3mr_action_scsiio to where
the error is detected. We cannot pass errors back from the
mpi3mr_enqueue_request to do this on a 'failed' mpi3mr in a centralized
place (since it has to be fire and forget).

Add comments about zero length SGLs never making it into
mpi3mr_prepare_sgls. Keep the code there for the moment, but we only set
cm->data to non-NULL when scsiio_req->DataLength is not zero. So the
datalength can't be zero and we can't send the zero SGLs.

Add commentts about other "impossible" tests in mpi3mr_prepare_sgls that
really should be simple asserts of some flavor.

Eliminate cm->error_code, since we can't pass data back from the
mpi3mr_prepare_sgl callback anymore.

In mpi3mr_map_request, call mpi3mr_enqueue_request for the no data case.
This seems to work even though we've not done the special zero length
handling that was in mpi3mr_prepare_sgls, giving further evidence to it
not actually being needed. This is needed for SCSI CDBs that have no
data to pass to the drive like TEST UNIT READY.

With this change, and the prior ones, we're now able to run with mpi3mr
on 128GB systems and very heavy disk load (so many buffers land > 4GB:
the driver instructs busdma to never use memory abouve 4GB, which may be
too conservative, but an issue for another time).

Sponsored by: Netflix