[E2E] Refactor the e2e testcases. (#4789)

### What this PR does / why we need it?
Refactor the e2e testcases.
- tests/e2e/multicard/test_weight_loader.py: Remove the unused code.
- tests/e2e/singlecard/multi-modal/test_internvl.py: Move to accuracy
test.
- tests/e2e/singlecard/test_aclgraph.py: Rename the file.
- tests/e2e/singlecard/test_embedding_aclgraph.py : Combine with
tests/e2e/singlecard/test_bge_model.py
- tests/e2e/singlecard/test_completion_with_prompt_embeds.py: Delete
eager mode and modify model to Qwen3-0.6B
- tests/e2e/singlecard/test_quantization.py: Modify model to
Qwen3-0.6B-W8A8
- tests/e2e/singlecard/test_vlm.py: Modify model to Qwen3-VL-8B

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

---------

Signed-off-by: menogrey <1299267905@qq.com>

tests/e2e/singlecard/pooling/test_embedding.py

@@ -24,7 +24,6 @@ from tests.e2e.utils import check_embeddings_close
 
 MODELS = [
     "Qwen/Qwen3-Embedding-0.6B",  # lasttoken
-    "BAAI/bge-small-en-v1.5",  # cls_token
     "intfloat/multilingual-e5-small"  # mean_tokens
 ]
 
@@ -57,3 +56,45 @@ def test_embed_models_correctness(model: str):
         name_1="vllm",
         tol=1e-2,
     )
+
+
+def test_bge_model_correctness():
+    queries = ['What is the capital of China?', 'Explain gravity']
+
+    model_name = snapshot_download("BAAI/bge-m3")
+    with VllmRunner(
+            model_name,
+            runner="pooling",
+            enforce_eager=False,
+    ) as vllm_aclgraph_runner:
+        vllm_aclgraph_outputs = vllm_aclgraph_runner.embed(queries)
+
+    with VllmRunner(
+            model_name,
+            runner="pooling",
+            enforce_eager=True,
+    ) as vllm_runner:
+        vllm_eager_outputs = vllm_runner.embed(queries)
+
+    with HfRunner(
+            model_name,
+            dtype="float32",
+            is_sentence_transformer=True,
+    ) as hf_runner:
+        hf_outputs = hf_runner.encode(queries)
+
+    check_embeddings_close(
+        embeddings_0_lst=hf_outputs,
+        embeddings_1_lst=vllm_eager_outputs,
+        name_0="hf",
+        name_1="vllm",
+        tol=1e-2,
+    )
+
+    check_embeddings_close(
+        embeddings_0_lst=vllm_eager_outputs,
+        embeddings_1_lst=vllm_aclgraph_outputs,
+        name_0="eager",
+        name_1="aclgraph",
+        tol=1e-2,
+    )