revert: remove sparse attention _linear inference-tensor clone hack

trellis/modules/sparse/attention/modules.py

@@ -74,21 +74,9 @@ class SparseMultiHeadAttention(nn.Module):
         if use_rope:
             self.rope = RotaryPositionEmbedder(channels)
 
-    @staticmethod
-    def _sparse_tensor_autograd_safe(x: SparseTensor) -> SparseTensor:
-        """If autograd is recording this forward, inference-mode feats break nn.Linear (cannot save for backward)."""
-        if not torch.is_grad_enabled():
-            return x
-        f = x.feats
-        is_inference = getattr(f, "is_inference", None)
-        if callable(is_inference) and is_inference():
-            return x.replace(f.clone())
-        return x
-
     @staticmethod
     def _linear(module: nn.Linear, x: Union[SparseTensor, torch.Tensor]) -> Union[SparseTensor, torch.Tensor]:
         if isinstance(x, SparseTensor):
-            x = SparseMultiHeadAttention._sparse_tensor_autograd_safe(x)
             return x.replace(module(x.feats))
         else:
             return module(x)