Kubernetes: fail fast if job pod was not scheduled (#3874)

After a job pod is created, wait and fail with `ComputeError` if the
pod has either not been scheduled or has already finished (probably
failed) within the scheduling timeout (10 seconds).

A new `watch` permission for `pods` in the namespace is required.

In addition, `run_job()` and `terminate_instance()` were refactored
to clean up objects on failures.

Part-of: #3871

Author: un-def

mkdocs/snippets/kubernetes/dstack-backend-role.yaml

@@ -6,7 +6,7 @@ metadata:
 rules:
   - apiGroups: [""]
     resources: ["pods"]
-    verbs: ["get", "create", "delete"]
+    verbs: ["get", "watch", "create", "delete"]
   - apiGroups: [""]
     resources: ["services"]
     verbs: ["get", "create", "delete"]