[WIP] move metadata to GPU.

nanovllm/kvcache/sparse/policy.py

@@ -88,6 +88,7 @@ class SparsePolicy(ABC):
         head_dim: int,
         num_cpu_blocks: int,
         dtype: torch.dtype,
+        device: torch.device = None,
     ) -> None:
         """
         Initialize policy resources.
@@ -102,6 +103,7 @@ class SparsePolicy(ABC):
             head_dim: Dimension per head
             num_cpu_blocks: Number of CPU blocks allocated
             dtype: Data type for tensors
+            device: Device for metadata storage (GPU recommended for performance)
         """
         pass