[Test] Add accuracy nightly test for new models (#4262)

### What this PR does / why we need it?
Add accuracy nightly test for new models：

PaddlePaddle/ERNIE-4.5-21B-A3B-PT
LLM-Research/Molmo-7B-D-0924
LLM-Research/gemma-2-9b-it
LLM-Research/gemma-3-4b-it
Shanghai_AI_Laboratory/internlm-7b
llava-hf/llava-1.5-7b-hf

- vLLM version: v0.11.2

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

tests/e2e/models/configs/Molmo-7B-D-0924.yaml

@@ -0,0 +1,13 @@
+model_name: "LLM-Research/Molmo-7B-D-0924"
+hardware: "Atlas A2 Series"
+model: "vllm-vlm"
+tasks:
+- name: "ceval-valid"
+  metrics:
+  - name: "acc,none"
+    value: 0.71
+max_model_len: 4096
+trust_remote_code: True
+apply_chat_template: False
+fewshot_as_multiturn: False
+gpu_memory_utilization: 0.8