Add a check for the proper size and type for the optimized scaled_index_add and index_select_cat, and fallback on pytorch implementation otherwise