add multi_stride for kernel shapes.

Add stride based multi-index similar to device functions. Between the device
gather and what's available for jit it looks like we're using lens instead of
strides to calculate indicies.

Seems to fix the 1d case of indices for this jit gather.

src/targets/gpu/kernels/include/migraphx/kernels/shape.hpp

@@ -128,6 +128,20 @@ struct shape
         result[0] = tidx;
         return result;
     }
+
+    /// Convert single index into a multi-index
+    constexpr index_array multi_stride(index_int idx) const
+    {
+        index_array result;
+        index_int tidx = idx;
+        for(diff_int is = result.size() - 1; is > 0; is--)
+        {
+            result[is] = tidx % strides[is];
+            tidx       = tidx / strides[is];
+        }
+        result[0] = tidx;
+        return result;
+    }
     /// Convert multi-index into a single index
     constexpr index_int single(index_array idx) const
     {