fix(python): Decode the k8s node selector and convert it as a dict. (#…

…3477)

## What do these changes do?

As titled.

During test, I find not all pods are scheduled to the same node.

```yaml
NAME                                             READY   STATUS    RESTARTS   AGE     IP            NODE           NOMINATED NODE   READINESS GATES
coordinator-ozgjbb-6c57549cf7-hhwqt              1/1     Running   0          4m2s    10.244.1.85   kind-worker    <none>           <none>
gs-engine-ozgjbb-0                               3/3     Running   0          3m58s   10.244.1.86   kind-worker    <none>           <none>
gs-engine-ozgjbb-1                               3/3     Running   0          114s    10.244.1.87   kind-worker    <none>           <none>
gs-engine-ozgjbb-2                               3/3     Running   0          108s    10.244.1.88   kind-worker    <none>           <none>
gs-engine-ozgjbb-ozgjbb-vineyard-etcd-0          1/1     Running   0          3m58s   10.244.3.71   kind-worker2   <none>           <none>
gs-interactive-frontend-ozgjbb-8d996bc8b-ctn6x   1/1     Running   0          3m58s   10.244.3.72   kind-worker2   <none>           <none>
```

We should support to add the node selector to the frontend and vineyard
etcd.

## Related issue number

Fixes #3476

---------

Signed-off-by: Ye Cao <[email protected]>

python/graphscope/deploy/kubernetes/resource_builder.py

@@ -126,7 +126,13 @@ def get_image_pull_secrets(image_pull_secrets):
 
     @staticmethod
     def get_node_selector(node_selector):
-        return node_selector
+        import base64
+        import json
+
+        decoded_node_selector = base64.b64decode(node_selector).decode(
+            "utf-8", errors="ignore"
+        )
+        return json.loads(decoded_node_selector)
 
     @staticmethod
     def get_user_defined_volumes(udf_volumes):