docs/install: add xCAT provisioner with rocky10/almalinux10 recipes - #2654

Adds xCAT to the Markdown recipe system, covering both of its provisioning
modes in a single guide. The mode is chosen at run time by ${xcat_mode} in
input.local rather than at document build time, so one guide and one
recipe.sh serve both, matching how the Warewulf recipe handles
provision-to-disk:

  stateless  compute nodes boot a RAM-resident image built in a chroot on
             the head node
  stateful   a base OS is installed to each node's local disk, then OpenHPC
             components are added to the running nodes

A "Select Provisioning Mode" section defines shell functions -- compute_exec,
compute_install, compute_group_install, compute_upgrade, compute_clean -- that
act on the image chroot when stateless and on the running nodes via xdsh when
stateful. The shared customization chapters call those helpers and are
otherwise untouched, so the InfiniBand, Omni-Path, memlock, GPU, syslog,
oneAPI and admin-tool sections are written once rather than once per mode.
Functions rather than variables because ${CHROOT} is only known after genimage
runs, and a function re-resolves it at call time.

Mode-specific steps live in templates/provisioner/xcat/ as stateless-* and
stateful-*, gated with ohpc_if so the generated script takes one path. Both
forked chapters group that work under a single heading -- "Stateless
Provisioning"/"Stateful Provisioning" and "Stateless Deployment"/"Stateful
Deployment" -- following the same aggregator-owns-the-grouping-header pattern
as "Cluster Admin Tools", so a reader follows one group and skips the other
outright. Nothing mode-specific sits outside those headings.

The deploy chapter starts Slurm on the head node first. Stateless nodes boot
afterwards and their slurmd reaches the controller on the first try, while the
already-running stateful nodes are then told to start theirs.

Details worth noting for review:

- The mode is a value, not a boolean. Every other input.local flag named
  enable_* turns an optional feature on or off, but this selects a mandatory
  mode, so it is spelled xcat_mode=stateless|stateful rather than a lone
  enable_stateful with no counterpart. It fails closed: anything else aborts,
  so a half-configured input.local cannot quietly netboot a fleet meant to
  install to disk.
- Image lookup refuses to guess. A head node holding images for several
  distributions or releases matches more than one, so the recipe stops unless
  exactly one matches or ${xcat_osimage} names it, and the chosen image is
  checked against the mode by its provmethod before anything is modified.
- The stateful wait requires every node to positively report a finished boot
  and to have lost a marker written to /run before the reset, so neither an
  unreachable node nor the system being replaced counts as ready. It gives up
  after ${install_timeout} rather than waiting forever on a failed install.
  It proves a reboot, not a successful install: a failed PXE attempt falling
  back to the existing disk would pass. Proving the install would need a
  per-run token written from the kickstart.
- The upgrade appendix spells out xdsh and --installroot commands per mode and
  re-derives ${CHROOT}, because the helper functions above exist only in the
  installing shell.

Both modes were deployed end to end on AlmaLinux 10.1 with xCAT 2.18.2 and two
compute nodes: the recipe ran start to finish in each, nodes provisioned to
tmpfs and to local disk respectively, Slurm showed both nodes idle, and a
two-rank MPI job ran across them.

Signed-off-by: Vinícius Ferrão <ferrao@versatushpc.com.br>