RDMA fixes, including memory deregistration

[ritvikrao]

Intended to fix issue #188. Reconverse was inadvertently overrwriting CMK_SRC_DEST_ACK to CMK_SRC_ACK, causing an error with destination ACKing. This fixes that (allowing the pingpong results to show for nodegroups) but isn't complete yet, there needs to be code to clean up RDMA buffers. currently this error happens at the end:

------- Processor 0 Exiting: Called CmiAbort ------
Reason: Unhandled C++ exception in user code: 0:55075:0:/Users/ritvik/charm_reconverse/reconverse-darwin-arm8/_deps/lci-src/src/network/ofi/backend_ofi.cpp:~ofi_device_impl_t:327<lci:Assert failed: false> err : Resource busy (/Users/ritvik/charm_reconverse/reconv
/Users/ritvik/charm_reconverse/reconverse-darwin-arm8/_deps/lci-src/lcrun: line 39: 55075 Segmentation fault: 11  "$@"
/Users/ritvik/charm_reconverse/reconverse-darwin-arm8/_deps/lci-src/lcrun: line 39: 55076 Segmentation fault: 11  "$@"

update: All RDMA tests now pass and pingpong runs correctly. This pr changes 2 things:

• adds CK_BUFFER_DEREG to charm-config.h, and other changes to allow Charm++ to use conv-rdma.h from reconverse, not old converse
• adds logic to deregister all RDMA buffers on exit

Using these changes requires the reconverse-rdma-fixes branch of Charm++; once this is merged into main, the reconverse-rdma-fixes branch will be merged into reconverse-specific-build in the charm repo

[lvkale]

Create a sort-of benign dependency on pup, but it is ok I think.