Add soft prompt tuning

[sdtblck]

• Adds a generalized insert_layer function to insert arbitrary layers in the model
• Uses it to add soft prompt tuning, but also could be used for other stuff (I'm thinking multimodal few shot training)

Should work in inference but haven't yet tested it.

We should be wary that soft prompt tuning / fine tuning will work better if we pad sequences up to max length, rather than packing them (which we do for pretraining).

I'll start working on a PR to simplify data loading + add an option to pad sequences, instead of pack.