Drop torchair (#4814)

aclgraph is stable and fast now. Let's drop torchair graph mode now.

TODO: some logic to adapt torchair should be cleaned up as well. We'll
do it in the following PR.

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

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
Co-authored-by: Mengqing Cao <cmq0113@163.com>

tests/e2e/nightly/models/test_deepseek_r1_w8a8_eplb.py

@@ -68,9 +68,6 @@ async def test_models(model: str) -> None:
         "cudagraph_mode": "FULL_DECODE_ONLY"
     }
     additional_config: dict[str, Any] = {
-        "torchair_graph_config": {
-            "enabled": True
-        },
         "enable_shared_expert_dp": False,
         "multistream_overlap_shared_expert": False,
         "dynamic_eplb": True,