Add support for CUDA

[sjpb]

• Adds a role cuda to install the NVIDIA CUDA toolkit onto hosts, see ansible/roles/cuda/README. This is not enabled by default, and requires OFED to be installed which currently must be done outside the appliance.
• Adds a new adhoc playbook ansible/adhoc/cudatests.yml which uses the cuda-samples utilities deviceQuery and bandwidthTest to test GPU functionality.
• Updates the stackhpc.openhpc role to:
  • Support defining gres parameter in openhpc_slurm_partitions parameter, to allow scheduling onto GPUs. (see v0.19.0)
  • Allow defining a custom openhpc package repo URL (see v0.19.0)
  • Allow multiple empty partitions (see v0.20.0)

Note that CUDA is not included in the default fat image. However if a compute host is added to the cuda group then the openstack.openhpc (fat image) or openstack.compute (compute image) will include the CUDA install.

[m-bull]

LGTM