Introduce query-wide execution state.

Currently impalads receive individual fragments that have their own self contained state.

However, these fragments have state that is duplicated for each fragment. We could benefit from having this shared state in a query wide state that all fragments refer to during execution.

This is the fundamental patch on top of which we need to implement a per query exec RPC, filter aggregation at different nodes across the cluster (not only the coordinator), per query reporting threads, etc.