fix(slurm): support CPU-only jobs on SLURM- #2137 (#2138)

## Summary

CPU-only jobs (`device_type: CPU`) submitted through the SLURM launcher
fail on CPU partitions:

- `common/distutils.py::setup` hardcodes `backend="nccl"` on the SLURM
branch, so a CPU job raises `ValueError: ProcessGroupNCCL is only
supported with GPUs, no GPUs found!`.
- `launchers/slurm_launch.py` sets `gpus_per_node = ranks_per_node`
unconditionally, so a CPU job still requests a GPU per node and SLURM
rejects it with `Requested node configuration is not available` on a CPU
partition.

The local-launch branch already honors `config["distributed_backend"]`
(which `map_job_config_to_dist_config` sets to `gloo` for
`DeviceType.CPU`). This brings the SLURM path in line and stops
requesting GPUs for CPU jobs.

## Changes

- `common/distutils.py`: SLURM branch uses
`config["distributed_backend"]` instead of a hardcoded `"nccl"` (matches
the local-launch branch).
- `launchers/slurm_launch.py`: request `gpus_per_node = 0` when
`device_type == DeviceType.CPU`.

## Test plan

Submitted a single-node CPU job (`device_type=CPU`, `qos=cpu_lowest`,
`cpus_per_task=192`) running a `Runner` that only writes files (no GPU
work).

- Before: rejected at submission (`Requested node configuration is not
available`), and when a GPU was not requested it failed at runtime with
the NCCL error above.
- After: the job initializes the Gloo backend (`Torch distributed
initialized with: file://...`), runs to `COMPLETED`, and produces all
expected outputs on a CPU node.

No change to GPU behavior: for non-CPU `device_type`, the backend
remains `nccl` and `gpus_per_node` is unchanged.

Author: lbluque

src/fairchem/core/common/distutils.py

@@ -121,7 +121,7 @@ def setup(config) -> None:
                 assign_device_for_local_rank(config["cpu"], local_rank)
 
                 dist.init_process_group(
-                    backend="nccl",
+                    backend=config["distributed_backend"],
                     init_method=init_method,
                     timeout=timeout,
                 )

src/fairchem/core/launchers/slurm_launch.py

@@ -278,7 +278,9 @@ def slurm_launch(cfg: DictConfig, log_dir: str) -> list:
         mem_gb=scheduler_cfg.slurm.mem_gb,
         timeout_min=scheduler_cfg.slurm.timeout_hr * 60,
         slurm_partition=scheduler_cfg.slurm.partition,
-        gpus_per_node=scheduler_cfg.ranks_per_node,
+        gpus_per_node=(
+            0 if cfg.job.device_type == DeviceType.CPU else scheduler_cfg.ranks_per_node
+        ),
         cpus_per_task=scheduler_cfg.slurm.cpus_per_task,
         tasks_per_node=scheduler_cfg.ranks_per_node,
         nodes=scheduler_cfg.num_nodes,