Realm: Refactor GPUIndirectXferDes by adding a scatter/gather gpu channel
This PR refactors GPUIndirectXferDes in the following way:
- Adds a separate channel
GPUScatterGatherChannel
for cuda gather/scatter operations which allows to "tweak" heuristics for the path planner ..e.g.frag_overhead
. This allows the path planner to choose this channel and therefore enable for example cuda gather operation into a GPU_FB on a remote node. In addition, the new channel allows to override thesupports_path
which allows to move all the decision making logic (on whether a gpu scatter/gather is possible) into the channel itself.
**PR does not contain any functional changes.
Edited by apryakhin