Implement multi-token prediction option for models

[zhaoyang-star]

From the Deepseek V3, it appears that models that predict multiple future tokens can exhibit significantly greater sample efficiency than models trained only on next-token prediction, plus the extra token heads can be used to implement speculative decoding to speed up inference (up to 3X in their experiments), without the need for a draft model.

It would be amazing to see multi-token prediction implemented in NeMo, as it would allow the community to easily experiment with this promising technique.