Support target masking (aka loss masking or label masking) for SFT datasets

[James Chen]

Right now, data loading and loss computation assume one is only doing LM pretraining, but it'd be useful to support packed SFT style datasets (i.e. datasets with cleanly delineated prompt/completion pairs, perhaps even a system prompt) and their corresponding masking.

I.e., the masks allow the attention module to reference the prompts/prefix, but only completions/targets' gradients are propogated.