Keep old gather/scatter implementation for COO_T

Author: dferre97

GNNlib/ext/GNNlibCUDAExt.jl

@@ -7,6 +7,14 @@ using GNNGraphs: GNNGraph, COO_T, SPARSE_T
 
 ###### PROPAGATE SPECIALIZATIONS ####################
 
+## COPY_XJ 
+
+## avoid the fast path on gpu until we have better cuda support
+function GNNlib.propagate(::typeof(copy_xj), g::GNNGraph{COO_T}, ::typeof(+),
+        xi, xj::AnyCuMatrix, e)
+    propagate((xi, xj, e) -> copy_xj(xi, xj, e), g, +, xi, xj, e)
+end
+
 ## E_MUL_XJ 
 
 ## avoid the fast path on gpu until we have better cuda support