ReFresh: Reducing Memory Access from Exploiting Stable Historical Embeddings for Graph Neural Network Training