[Core][MM] Use non-blocking CPU-GPU copy of multimodal data (#28141)

Signed-off-by: Lukas Geiger <lukas.geiger94@gmail.com>

vllm/multimodal/utils.py

@@ -444,7 +444,9 @@ def group_mm_kwargs_by_modality(
 
             if device is not None:
                 mm_kwargs_group = json_map_leaves(
-                    lambda x: x.to(device=device) if isinstance(x, torch.Tensor) else x,
+                    lambda x: x.to(device=device, non_blocking=True)
+                    if isinstance(x, torch.Tensor)
+                    else x,
                     mm_kwargs_group,
                 )
         else: