🚀 Feature

Currently, start_method for xmp.spawn method is set to "fork" only for Colab env. On Kaggle it has a considerable performance impact. Idea is to make start_method configurable for XLA and set as "fork" by default for TPUs