btl/uct using wrong topology

[rhc54]

In mca_btl_uct_component_open, a decision on the number of contexts to use is based on getting the number of cores on the node. However, HWLOC will return that value based on the number of cores available to the PRRTE daemon - which is not the same as the number of cores available to the application. For example, the user may have specified a cpuset (a.k.a., "soft" cgroup) that only applies to the application's procs.

Correct computation of the number of available cores requires that you get the PMIX_JOB_CPUSET, mask the complete_cpuset against it to find the actually available cpus, and then obtain the count of cores from within that cpuset.

[gpaulsen]

Removing blocker label as btl/uct is low priority.