[FEA] Compute the desired number of GPU pytest workers from the available GPU memory

Update the relevant CI scripts to implement this proposal. We can probably augment it to run on the GPU with the most available memory.

```bash
# Use memory of device 0 to determine the number of pytest-xdist workers
gpu_mem_mib=$(nvidia-smi --query-gpu=memory.total --format=csv,noheader,nounits | head -n1)
gpu_mem_gb=$((gpu_mem_mib / 1024))
mem_gb_per_worker=3
num_workers=$((gpu_mem_gb / mem_gb_per_worker))
if [ $num_workers -lt 1 ]; then
num_workers=1
fi
```

_Originally posted by @bdice in https://github.com/rapidsai/cudf/pull/19355#discussion_r2200851267_

We can have the script return the device is and the number of purest workers