Dev/rackscale shmem allocators

[hunhoffe]

This PR completes the process of using memory with NUMA affinity for rackscale benchmarks. Particularly, this includes:

• Allocating shmem RPC buffers in client affinity shmem
• Allowing clients and the controller to allocate memory from any affinity by creating a unique affinity NodeId for 1 shmem region per kernel. This allowed the client frame mapping state to be deleted, as that state is now embedded in the node id. This feature will be needed for dynamic replication
• Associate the ShmemAllocator used for ProcLog replicas with a shmem affinity - this can be used in the future when a replica may be used remotely (in the case of dynamic replication)
• Specify a host physical numa node to assign to clients' non-shmem memory
• Change rackscale benchmarks to run with 1, 2, 3, and 4 clients (on 4 socket machines) - this is becuase 4 clients may not be good to measure as one client will be co-located on the same NUMA node as the controller.
• Fix a bug in rackscale baseline benchmark tests

Of note, log replicas were already allocated using local shmem, so no changes were made there.

[hunhoffe]

Found a bug. Will commit a fix shortly and wait to merge until after.

[hunhoffe]

Should be ready to go again