When scheduling dynamic executors, order nodes by usage instead of picking a node at random

[Alexis-D]

The current logic just picks a random node on which the executor will fit:

https://github.com/palantir/k8s-spark-scheduler/blob/36236cf049a5c730afa5e7107eb3ef0fb624cc35/internal/extender/resource.go#L435-L442

This PR simply tries to fit the new executor in the 'fullest' node.

[svc-autorelease]

Released 0.43.0