LU-7734 lnet: cpt locking

When source nid is specified it is necessary to also
use the destination nid. Otherwise bulk transfer will end up
on a different interface than the nearest interface to the
memory. This has significant performance impact on NUMA
systems such as the SGI UV.

The CPT which the MD describing the bulk buffers belongs to
is not the same CPT of the actual pages of memory.
Therefore, it is necessary to communicate the CPT of the pages
to LNet, in order for LNet to select the nearest interface.

The MD which describes the pages of memory gets attached to
an ME, to be matched later on. The MD which describes the
message to be sent is different and this patch adds the
handle of the bulk MD into the MD which ends up being
accessible by lnet_select_pathway(). In that function
a new API, lnet_cpt_of_md_page(), is called which returns the
CPT of the buffers used for the bulk transfer.
lnet_select_pathway() proceeds to use this CPT to select
the nearest interface.

Signed-off-by: Amir Shehata <amir.shehata@intel.com>
Change-Id: I4117ef912835f16dcdcaafb70703f92d74053b9b
Reviewed-on: https://review.whamcloud.com/24085