Replace round-robin front-end node selection with least-loaded algorithm.

No changes to logic

Fix the limit at which slow suspend/resumes are logged

Array of active job suspend IDs was using -1 to indicated empty slots in the array, but the array was of type uint32_t. Cray might have noted a problem in this.

Increase the number of simultaneous job suspends or resumes that
slurmd can handle from 8 to 64.
See bug 826

Increase MAX_THREAD from 130 to 256

If specialized cores are allocated for a job and that job request
CPU binding using a cpu_map, the cpu_map was being interpretted
as a cpu_mask and a cpu_map of zero was being treated as invalid.
This is now fixed.

This reverts commit 859839a7
The ntasks_per_core option was previously really treated like number of CPUs (rather than tasks) to allocate per core, which seems to be what is desired.

This needs a lot of work, but checking in the framework for now.

wait for.

based on the mask given.

task/affinity.

thread in a core.

it can bind cyclically across sockets.

0c090f95

add a PriorityFlags option of CALCULATE_RUNNING. If set, then the
priority of running jobs will continue to be recalculated
periodically. The PriorityFlags value reported by sview and
"scontrol show config" will be reported as a string rather
than its numeric value.

Previous logic assumed cpus_per_task=1, so ntasks_per_core option
could spread the job across more cores than desired

cpus-per-task support: Try to pack all CPUs of each tasks onto one socket.
Previous logic could spread the tasks CPUs across multiple sockets.

Previous logic was counting CPUs, but assuming each task would
only use one CPU.

This reverts commit b22268d8.