`chat_dataset(...)` should use newer `masking_strategy` arg

[vadimkantorov]

It is currently using the deprecated train_on_input

[pbontrager]

All of the datasets still have train_on_input. This task requires updating:

• update all datasets to use masking_strategy (Chat, Instruct, all named datasets)
• git grep train_on_input to update all the docs and tests and configs (only 3)
• remove train_on_input from Message and datasets

@ebsmothers @joecummings is this the behavior we want?

[RdoubleA]

I had an early version of the full deprecation in #2533 but I can't take it forward. It just needs to be rebased and cleaned up

[ysurs]

I will start working on this issue @joecummings, I will check the above PR first to get an understanding, will ping you on discord if i have any follow ups.