[CI][Nightly] Support local debugging for multi-node CI test cases (#4489)

### What this PR does / why we need it?
 This patch mainly doing the following things:
1. Make k8s/lws optional for multi-node testing, allowing developers to
run multi-node tests locally by actively passing in the IP addresses of
all nodes.
2. Allows passing a custom proxy script path in the config file to load
the proxy.

- vLLM version: v0.11.2

---------

Signed-off-by: wangli <wangli858794774@gmail.com>

tests/e2e/nightly/multi_node/test_multi_node.py

@@ -7,8 +7,8 @@ from modelscope import snapshot_download  # type: ignore
 from requests.exceptions import ConnectionError, HTTPError, Timeout
 
 from tests.e2e.conftest import RemoteOpenAIServer
-from tests.e2e.nightly.multi_node.config.multi_node_config import (
-    DISAGGREGATED_PREFILL_PROXY_SCRIPT, MultiNodeConfig)
+from tests.e2e.nightly.multi_node.config.multi_node_config import \
+    MultiNodeConfig
 from tools.aisbench import run_aisbench_cases
 
 prompts = [
@@ -100,8 +100,10 @@ async def test_multi_node() -> None:
     disaggregated_prefill = config.disaggregated_prefill
     server_port = config.server_port
     proxy_port = config.proxy_port
-    server_host = config.cluster_ips[0]
-    with config.launch_server_proxy(DISAGGREGATED_PREFILL_PROXY_SCRIPT):
+    server_host = config.node_info.ip
+    proxy_script = config.envs.get("DISAGGREGATED_PREFILL_PROXY_SCRIPT", \
+        'examples/disaggregated_prefill_v1/load_balance_proxy_server_example.py')
+    with config.launch_server_proxy(proxy_script):
         with RemoteOpenAIServer(
                 model=local_model_path,
                 vllm_serve_args=config.server_cmd,