Regarding #1, there is not scheduling once execution reaches the k8s cluster, but the scheduling is done before that, using the glue.
I think that in the paid tier there is support for dynamic GPU scheduling that allows you to specify allocation requirements on the task.