Refactor launchers to add ray cluster mode

[rayg1234]

• This add a new mode to launch Ray jobs with fairchem: use_ray and RayClusterConfig
• Refactor and simplifies the _cli so that the main launching logic is separated and moved to the launchers module

Previously we could only launch slurm jobs and local jobs with torch elastic, with adding Ray, we have the following configuration options:

The ray cluster is adapted from the fairray script. To launch a job, we start a head job and a worker job

• The head node runs the Ray gcs server and the dashboard etc. It also runs the driver program
• The worker nodes to just supply the resources for Ray
  If the job requests N nodes, then 1 node would be allocated for the head and N-1 nodes allocated to the worker job

We can launch jobs the same way:
fairchem -c ray_job.yaml

Added an example config (uma_ray_demo) to train uma using Ray:

fairchem -c configs/uma/training_release/uma_ray_demo.yaml cluster=h100 dataset=uma_debug job.scheduler.slurm.qos=h100_ocp_high job.scheduler.num_nodes=2 max_atoms=100

this is equivalent to running it on slurm via:

fairchem -c configs/uma/training_release/uma_sm_direct_pretrain.yaml cluster=h100 dataset=uma_debug job.scheduler.slurm.qos=h100_ocp_high job.scheduler.num_nodes=2 max_atoms=100

This should not affect any previous jobs or configs.

[misko]

LGTM!

[lbluque]

lgtm - minor comments and nits.

[lbluque]

nit, for local and slurm (use_ray=False), we use Submitit either calling it directly or using an executor for slurm. But when reading the code this is a bit unclear - took me a bit to piece it back together. Would be great to create an explicit local_launch module too (which would be pretty empty, just have the runner_wrapper in there).