Parallel fixes - #2585

kwargs['INDEX1'] was computed as str(i+i) (i.e. 2*i) instead of the intended
1-based batch-slot index str(i+1). Any code-interface command using the
%INDEX1% placeholder received 0,2,4,... for slots 0,1,2,..., silently
selecting the wrong node files / bindings for batch slots >= 1.

- The Dask branch of __shutdownParallel guarded on rayInstanciatedOutside
  instead of daskInstanciatedOutside; with an externally provided scheduler
  RAVEN could tear down resources it does not own.
- When RAVEN owns the cluster, call Client.shutdown() so the scheduler retires
  ALL workers (including the ssh-launched remote 'dask worker' processes,
  which previously had no teardown path and were left as zombies), then
  terminate the tracked head-node worker and scheduler subprocesses.
- Track the head-node dask worker Popen (previously a dropped local variable).
- Close ray_head.ip and dask worker log file handles via context managers.
- Guard Ray address_info debug prints with .get(); 'redis_address' was
  removed in Ray 2.x and raised KeyError under verbose debugging.

RayRunner.kill() and DaskRunner.kill() only dropped the local ObjectRef /
Future, so 'killed' remote tasks kept executing and consuming cluster cores
(affecting optimizer early stopping, terminateJobs and terminateAll). Now
kill() issues ray.cancel(force=True, recursive=True) / Future.cancel() before
releasing the reference, with a warning if cancellation fails.

…; propagate tracebacks (B7, quick-ref #7)

- InternalRunner gains an explicit tri-state runSucceeded flag and a
  failureInfo string (formatted traceback / RayTaskError text), exposed via
  getFailureInfo(). getEvaluation() now returns Error() based on the explicit
  flag; the 'None return == failure' sentinel remains only as a legacy
  fallback when no outcome was recorded.
- SharedMemoryRunner: the bare lambda thread target is replaced by a wrapper
  that catches exceptions, records the traceback and the outcome; functions
  that legitimately return None are no longer marked failed.
- RayRunner/DaskRunner: success and failure are recorded explicitly on
  ray.get()/Future.result(); failure text is kept. DaskRunner no longer
  re-raises inside _collectRunnerResponse, so getEvaluation() cannot blow up
  the collecting thread (getReturnCode retains its failure semantics).
- JobHandler.__checkAndRemoveFinished logs the failure details and stores them
  under 'failureInfo' in the failed-job metadata.

PyThreadState_SetAsyncExc is ignored while a thread is blocked in C
extensions or system calls, so the previous 'while alive: sleep; kill()' loop
could spin forever. kill() now retries for a bounded 10 s window, then warns
and abandons the (daemon) thread, marking the run as failed explicitly.

…s, quote variables (B9)

- start_remote_servers.sh: '--help' used 'return' which is invalid in an
  executed script (now 'exit 0'); error paths exit non-zero; START_OUTPUT was
  referenced by tee but never defined (now <output>_start.log); the parsed
  --num-gpus value was silently dropped and is now forwarded to start_ray.sh;
  variable expansions are quoted.
- start_ray.sh: new NUM_GPUS positional argument, passed to 'ray start
  --num-gpus' only when non-negative; optional bash profile; quoted variables.
- start_dask.sh: the remote bash profile is now optional (previously
  'source ""' failed when unset); 'cd' is checked; quoted variables.

…le (B10)

PBS and MPI-legacy modes read os.environ[<nodefileenv>] directly, producing a
raw KeyError traceback when the variable is undefined. Both now raise a RAVEN
IOError naming the missing variable.

…, add unit tests (quick-ref #8)

- New ClusterMode.ClusterSimulationMode base class holds the previously
  triplicated modifyInfo logic: node-file reading, batch-size clamping,
  per-batch node_%INDEX% file splitting, mpiexec precommand assembly and
  thread postcommand handling. Slurm and PBS modes are now thin subclasses
  that only perform scheduler-specific node discovery and submission (the MPI
  legacy mode is intentionally left untouched).
- ClusterUtils gains pure, framework-independent functions (readNodeFile,
  computeBatchSize, writeNodeSubFiles, buildMPIPrecommand, sanitizeJobName)
  used by the base class.
- Job-name validation is unified via ClusterUtils.sanitizeJobName: both modes
  now accept alphanumerics, '_' and '-' (the Slurm mode previously rejected
  '-' while PBS allowed it), PBS keeps its 15-character truncation, and both
  raise RAVEN errors instead of bare IOError/print.
- Empty PBS node files now raise a clear error (previously silently produced
  batchSize=1); PBS node-file handles are context-managed via readNodeFile.
- New standalone unit tests (21 cases) in CustomModes/tests/, runnable
  without a built RAVEN: python3 ravenframework/CustomModes/tests/test_cluster_utils.py

New optional <useSrun/> mode option: run precommands become
'srun --overlap --exact -n <NumMPI> [MPIParam...]' instead of
'mpiexec <NodeParameter> <nodefile> -n <NumMPI>'. Slurm assigns per-step
resources itself, so the per-batch node_%INDEX% files are skipped entirely,
removing the hostfile-flag dependence on the MPI flavor (MPICH -f vs OpenMPI
--hostfile) and working with all major MPI stacks via PMI/PMIx. Node
discovery is kept so runInfo['Nodes'] remains available for internal-parallel
cluster bring-up. Extra <MPIParam> entries are passed to srun (e.g.
--mpi=pmix). Includes unit tests for the precommand construction.

…ref #10)

New <daskJobqueue> RunInfo element (used with <internalParallel>True
</internalParallel> and <parallelMethod>dask</parallelMethod>): the element
text selects the scheduler ('slurm' or 'pbs') and the XML attributes are
passed to the dask_jobqueue SLURMCluster/PBSCluster constructor. Example:

  <daskJobqueue memory="4GB" queue="short" account="proj1">slurm</daskJobqueue>

Dask workers are then submitted AS scheduler jobs by dask_jobqueue, removing
the inter-node-ssh requirement and the hand-rolled bring-up scripts for this
path, with clean teardown (client + cluster close cancels the worker jobs).
'memory' is required (dask_jobqueue mandates it); 'cores' defaults to
numProcByRun, 'jobs' (number of worker scheduler jobs) defaults to batchSize,
'walltime' defaults to <expectedTime>; other attributes pass through
verbatim. Fails with a clear error when dask_jobqueue is not installed.
Kwargs assembly lives in ClusterUtils.assembleDaskJobqueueKwargs (pure,
covered by unit tests). The cluster object is excluded from pickling.