Update data masking for agnostic pretraining data

[aakankshaduggal]

Update data masking for agnostic pretraining data

adding @Maxusmusti and @RobotSail for more details around the requirements from the training

[bbrowning]

The goal here is for SDG to be able to generate data once that's usable with any student model. Today, the pretraining samples we generate have to be generated in a format that matches the chat prompt of the intended student model used for training. Instead, we want to generate data with SDG once and have training do whatever post-processing it needs to our pretrain samples to adapt them to the student model being trained.