[CI] refect e2e test (#4799)

### What this PR does / why we need it?
This PR updates the CI configuration and adjusts a set of end-to-end
(e2e) tests under tests/e2e/multicard, in order to refactor the test
suite and ensure compatibility with current codebase and CI workflows.

1. tests/e2e/multicard/test_prefix_caching.py: change model to Qwen3-8B
and rename the test case
2. tests/e2e/multicard/test_quantization.py: rename the test case
3. tests/e2e/multicard/test_qwen3_moe.py: remove duplicate test and
rename test cases
4. tests/e2e/multicard/test_qwen3_next.py: rename test cases and change
the W8A8 pruning model to the W8A8 model and remove the eager parameter
5. tests/e2e/multicard/test_shared_expert_dp.py: rename test case and
remove the eager parameter
6. tests/e2e/multicard/test_single_request_aclgraph.py: rename test case
and change Qwen3-30B to Qwen3-0.6B
7. tests/e2e/multicard/test_torchair_graph_mode.py: delete test cases
about torchair

- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

Signed-off-by: hfadzxy <starmoon_zhang@163.com>

tests/e2e/multicard/test_quantization.py

@@ -25,16 +25,17 @@ from modelscope import snapshot_download  # type: ignore
 from tests.e2e.conftest import VllmRunner
 
 
-def test_models_distributed_quantized_W8A8():
+def test_qwen2_5_w8a8_external_quantized_tp2():
     example_prompts = [
         "The president of the United States is",
     ]
     max_tokens = 5
-    with VllmRunner(snapshot_download("neuralmagic/Qwen2.5-3B-quantized.w8a8"),
-                    tensor_parallel_size=2,
-                    max_model_len=4096,
-                    gpu_memory_utilization=0.8,
-                    enforce_eager=False) as vllm_model:
+    with VllmRunner(
+            snapshot_download("neuralmagic/Qwen2.5-3B-quantized.w8a8"),
+            tensor_parallel_size=2,
+            max_model_len=4096,
+            gpu_memory_utilization=0.8,
+    ) as vllm_model:
         vllm_output = vllm_model.generate_greedy(example_prompts, max_tokens)
 
     golden_results = [