It isn’t recommended to run multiple head nodes on a single machine, but if you really need to, I think you could use kill -9 to directly kill a specific GCS server process. You can use pgrep gcs_server to find these instances, and if you kill one, the associated Ray node should also clean itself up.
I believe ray.init() will autoconnect to the latest Ray cluster, so you may also have to pass in the specific IP printed by the ray start command for this to work.