The federation needs to make a job_desc when requeueing jobs to
siblings.

Since a persistent connection can only be established by SlurmUser this
prevents non-slurmuser users from calling the rpcs. It also requires that all
slurmctlds in the federation have the same SlurmUser.

If the job can't start now, just submit the job to all siblings.

_update_sibling_job_siblings()

like it does in slurm_send_recv_msg. The resp needs to be inited before
_check_send it called.

Sibling jobs have to get lock from the origin cluster in order to attempt to
allocate nodes. If it gets the allocation then it lets the origin cluster know
and the origin cluster will set the siblings jobs, if any, into a REVOKED state
and purge the jobs. If the sibling job is the only sibling then it assumes the
lock and attempts to start the job to avoid extra communications. If nodes can't
be allocated then the job releases the lock for another cluster to try.

for fed sibling jobs that don't start.

To handle JOB_REVOKED

Overwriting _size leads to the bitmap being the wrong length.

CID 160063.

The safe_unpackstr_xmalloc() call could jump to unpack_error,
which would leak memory allocated for bitmap. Allocate only after
the unpackstr has succeeded.

Coverity 160092 (+ more due to macro expansion leading to repeats).

Code would return NO_VAL if the requested frequency was greater than
the highest available.

While here improve the errors printed to the slurmstepd log location,
and change the initial check against nfreq to test for zero. (Rather
than (uint8_t) NO_VAL which it could never be set to.)

Bug 3335.

Bug 2508

If job is allocated nodes which are powered down, then reset job start time
    when the nodes are ready and do not charge the job for power up time.
bug 3411

No changes in logic

Modify get_my_user_name to return a FAILURE if the user_name cannot
be determined, as was handled inconsistently.

the function.

Bug 3312 and 3332

jobid and stepid to be used if needing to kill job because it hadn't started yet.

Add logic to monitor Uncorrectable Memory Errors (UME) and notify
  active jobs in case they run for a while afterwards. This copies
  logic from knl_generic to knl_cray. There may be a different UME
  monitoring system for Cray systems in the future. The original
  knl_generic development is in commit 56ff27da
bug 3341