NUMA-awareness for separate join build

Currently joins are executed by a single thread, so generally the memory allocations will come from the local NUMA node (assuming the thread stays on the same NUMA node).

With the separate join build on a NUMA machine, some probe threads will likely be on a different NUMA node from the build, resulting in inefficiency. This may not be a major problem because of prefetching, etc, but filing this JIRA to track the potential issue.