remove cpu offload

It's not used in most benchmarks. And when it is, we don't have to move the pipeline to `cuda`.

sayak_lcm_benchmark.py

@@ -22,7 +22,6 @@ def load_pipeline(standard_sdxl=False):
         pipe.load_lora_weights(LORA_ID)
 
     pipe.to(device="cuda", dtype=torch.float16)
-    pipe.enable_model_cpu_offload()
     return pipe