you kill, not cancel, actors (stanford-crfm#871)

src/levanter/infra/ray_tpu.py

@@ -172,7 +172,7 @@ def do_run(self, remote_fn, coordinator_ip, slice_id, num_slices) -> _TpuRunResu
         logger.exception(e)
         for actor in actors:
             try:
-                ray.cancel(actor)
+                ray.kill(actor)
             except Exception:
                 logger.exception("Failed to kill actor after primary failure")
         return futures