Only support the contiguous case for cublas matmul.

src/tensor.rs

@@ -275,6 +275,15 @@ impl Tensor {
                 op: "matmul",
             });
         }
+        if let crate::DeviceLocation::Cuda { .. } = self.device().location() {
+            if !self.is_contiguous() || !rhs.is_contiguous() {
+                // It looks like the cublas implementation of XgemmStridedBatched only supports
+                // non-standard strides on the batch dimension.
+                return Err(Error::RequiresContiguous {
+                    op: "matmul-cublas",
+                });
+            }
+        }
 
         let m = a_dims[dim - 2];
         let k = a_dims[dim - 1];