[Transform] SpinQuant fix OOM (vllm-project#1976)

SUMMARY:
"When using SpinQuantModifier for some fuse operations, it is necessary
to add the torch.no_grad decorator. Otherwise, PyTorch will capture the
grad graph by default, leading to a gradual increase in memory usage. I
encountered a CUDA OOM issue when rotating the MOE model, and the OOM
problem was resolved after fixing it."


TEST PLAN:
"Performed code quality evaluation locally"

Signed-off-by: LeiZhang <isleizhang@outlook.com>

Author: zhanglei1172

src/llmcompressor/modifiers/transform/spinquant/base.py

@@ -144,6 +144,7 @@ def on_initialize(self, state: State, **kwargs) -> bool:
 
         return True
 
+    @torch.no_grad()
     def on_start(self, state: State, event: Event, **kwargs):
         self.started_ = True