Fix for CR_SOCKET_MEMORY and per-CPU memory limti

The logic in select/cons_res had a bad calculation where it used
the CPUs per socket rather than CPUs per core in determining how
much memory a job would require (e.g. 6 cores per socket, 2 threads
per core; if a job requested 4 tasks, the algorithms was determining
the memory requirement would be that of 48 CPUs [4*6*2] rather than
8 CPUs [4*2])