using --cpu-only + --gpu-only options supported by ReFrame

[boegel]

I had this bit written up for the docs on the EESSI test suite, but neither of these options doesn't seem to work as intended with our GROMACS/TensorFlow tests:

#### Filtering by device (CPU, GPU)

By default, ReFrame will generate variants of tests for each applicable
device type, based on the specified [`features`](https://reframe-hpc.readthedocs.io/en/stable/config_reference.html#config.systems.partitions.features) for system partitions (in the ReFrame configuration file) and [`valid_systems`](https://reframe-hpc.readthedocs.io/en/stable/regression_test_api.html#reframe.core.pipeline.RegressionTest.valid_systems) value of the available tests.

To only run checks on CPU, you can use the [`--cpu-only` option](https://reframe-hpc.readthedocs.io/en/stable/manpage.html#cmdoption-cpu-only).

To only run tests on GPU, you can use the [`--gpu-only` option](https://reframe-hpc.readthedocs.io/en/stable/manpage.html#cmdoption-gpu-only).

For example, to only run tests on GPU:

```
reframe --gpu-only
```

You can use `--list` to check the impact of these options on generated checks.

• When using --gpu-only --list, no checks are generated
• When using --cpu-only --list, I see GROMACS CUDA modules being used to generate tests (they should only be used for GPU checks?)

[smoors]

--gpu-only filters out tests that do not have gpus_per_node defined.
--cpu-only filters out tests that have gpus_per_node defined.

in the GROMACS test, gpus_per_node is defined after the setup stage (hook assign_one_task_per_compute_unit), which is after the filtering, thus all/no tests are filtered out with --gpu-only and --cpu-only, respectively.

note that in ReFrame 4.4 those options are deprecated; there is a new option -E which allows filtering on any variable or parameter.

[smoors]

the reason we set the resources after the setup phase is that we can only know the available resources for concrete test cases (via the partition).