Only checks that device is one of the recognized keywords – "cpu",
"cuda", "mps", or "gpu"/"auto" – so a typo is caught before the
simulator runs, the same reason check_train_controls() runs early. This
is the whole check for "cpu". Turning "cuda"/"mps"/"gpu"/"auto"
into a concrete, available device needs torch loaded, which
density_estimator = "linear_gaussian" never requires; that step is
resolve_device()'s job, called only once a torch-backed estimator is
about to train.