Recipes for Sequential Pre-training of Multilingual Encoder and Seq2Seq Models