Fix: Ensure peer cache management supports all stacks

This adds a new function crm_get_cluster_proc() to return the process flag
corresponding to the cluster's messaging layer, so that node status can
be reliably determined regardless of stack (previously the code was
inconsistent about checking for all stacks).

Pacemakerd uses the new function when initializing the process list
broadcast to other nodes. Without this, each broadcast turns off
the flag in other nodes' peer caches, preventing those nodes from recognizing
the state change if this node is actually lost later.