Partitial fix for new job array data structure

Fix squeue --array/-r option to print individual job array task
records even if then are all in a single job record from slurmctld.

Change un/pack_bit_str() to pack_bit_str_hex(), which has much
better scalability and performance.

set exit code on a failure mode previously ignored

Move the formatting of a job array's task ID string from slurmctld
to the client to reduce system overhead. There is logic in place
to translate the full bitmap into a string, but in the worst case
this can take roughly one minute (e.g. 1-1000000:2,2) so by default
only the first 1024 bytes in the string are set. There is logic in
the code to translate the full string, but commented out.

Fix some anomalies in the main scheduling loop for job arrays
Improve logging in the main scheduling loop for job arrays

Now able to run job arrays based upon the new job array data structure.

This adds logic to copy a job record for a job array when we actually
schedule one of them. More work is still needed.

Add logic to build job array task ID string with a step function
rather than using the bit_unfmt() function, which is terribly slow
under those conditions for large job arrays.

Modify the job state RPC, squeue, sview, and scontrol to display
the new job array task ID string.

Add new data structure for job arrays so only one job record is
required for all tasks in a pending job array
Initialize and free the new structure

Add squeue support for SQUEUE_FORMAT2 environment variable, which
is equivalent to the -O/--Format option.

Report correct partition associated with a step if the job is submitted to
multiple partitions. Previously all of the partitions that the job could
run in would be reported for the step.

Added squeue -O/--Format option that makes all job and step fields available
for printing.

Add the uid cache functions thread safe
add function to clear the uid cache when checking for memory leaks

squeue has to convert uid_t to a char* username a lot. Every line of
output(with the default format) and for every comparison when sorting by
username. This causes multiple system calls to open and read /etc/passwd
on standard systems and on my system running YP there were 17
unnecessary entries in strace per line in the output.

The cache is very simple, just a sorted array that is resorted for every
new name. The number of unique usernames is likely going to be so low
that any machine can handle it. No attempt is made at freeing memory for
the same reason.

Add more detail to sched/backfill logging

Conflicts:
	testsuite/expect/test21.30

remove crufty(?) log message

If MinJobAge is configured too low, then job records can be purged
before these tests issue the job requeue test, resulting in their
failure

Conflicts:
	src/plugins/select/cons_res/job_test.c

Without this, the record of gres allocated to steps will remain
allocated and not be available to job steps after being requeued.