assorted changes for Cray system use of proctrack/sgi_job

This adds a link test for libjob.so
 * salloc needs direct support (Makefile.am),
 * however, X_AC_SGI_JOB comes much later in configure.ac.

(An alternative would be, since the libjob interface has practically not
 changed for 2.6 kernels, to integrate its 19 ioctls into slurm.)

This adds detection and use of SGI process aggregate job container IDs for
salloc interactive sessions.

The preferred and documented way to support this on a Cray system is by
enabling the provided pam_job.so via /etc/pam.d/common-session.

There is a header dependency on job.h. This file depends on the optional
cray-libjob-devel package, which installs into /opt/cray/job/<version>.

This package is however not always installed or may not be up-to-date.
Hence the patch "cheats" by duplicating the known prototype of job_getjid().

To work properly, select/cray requires proctrack/sgi_job. In fact, due to the
way the container functions are called by the slurmstepd, it will only work
properly with this plugin.

I have considered alternatives, such as falling back to using the SID to
confirm the allocation. But this attempt to support configuration errors
creates other problems, such as less stringent error checking.

This equips the slurmstepd to use the SGI process aggregate container
ID to confirm the ALPS reservation.

The way it is coded allows ALPS to return a temporary error in confirming the
reservation once, the job will then be requeued.

On the other hand, it is mandatory that the Cray job service works correctly,
therefore errors are returned if
 * job container creation fails or
 * the job is not attached to a container (anticipating later failure in
   _fork_all_tasks() when slurm_container_add() will fail for the same reason).

The patch relies on the internals of the proctrack/sgi_job plugin in order to
avoid duplicating code. This dependency is made explicit by a configuration
check of a subsequent patch.

With these two pieces in place, the frontends are set to DRAINING if a system
administrator forgets to enable the /etc/init.d/job service, as shown in the
following log entries:

slurmd log:
[2011-04-26T14:56:04] Launching batch job 134 for UID 21215
[2011-04-26T14:56:04] [134] no PAGG ID: job service disabled on this host?
[2011-04-26T14:56:04] [134] could not confirm ALPS resId 253
[2011-04-26T14:56:04] [134] job_manager exiting abnormally, rc = 4014

slurmctld log:
[2011-04-26T14:56:03] ALPS RESERVATION #253, JobId 134: BASIL -n 2 -N 1 -d 1 -m 16000
[2011-04-26T14:56:03] sched: Allocate JobId=134 NodeList=nid000[16-17] #CPUs=24
[2011-04-26T14:56:04] error: slurmd error 4014 running JobId=134 on \
			front_end=gele2: Slurmd could not set up environment for batch job
[2011-04-26T14:56:04] update_front_end: set state of gele2 to DRAINING
[2011-04-26T14:56:04] completing job 134

This uses the SGI container process aggregate ID to confirm the job reservation.

It falls back to using the alloc_sid in case of failure. This fallback
should be considered really only as a last resort, since it is known that
session IDs are not unique across multiple login nodes and therefore the
confirmation of ALPS reservations will fail whenever there is a SID collision
(the likelihood increases with system size).

This extends the Cray-specific select_jobinfo struct with a confirmation cookie
field, which is to be used by later patches to store the session SID or PAGG
container ID.

There is a slight incompatibility with regard to pack/unpack, due to the new
confirm_cookie format. Since not many Cray installations exist yet, I would
like to suggest to not do an extra bump of the API version.

This is just for consistency with other proctrack plugins, which all return 0
to indicate "not found", rather than (uint64_t)-1.

The select/cray plugin discovers the topology as part of its initialisation and
generates a node ranking. No further topology information is required by the
plugin, hence this patch sets the default TopologyPlugin to topology/none.

recent changes for Cray system support

to create job allocation.

    --mem-per-cpu options on a heterogeneous cluster. Patch from Bjorn-Helge
    Mevik, University of Oslo.

have slurm srun command installed on a cray system instead of the wrapper.

to function as desired.