More APIs for high-level library development

Hi,

I have been developing a [high-level library](https://github.com/ZhiyuanChen/DanLing/tree/develop) for PyTorch and Accelerate for quite some time.
The reason why I'm doing it is because I'm tooooo lazy to write all the logging and IO stuff, but I still want full control with the training loop (nobody tells me how to train my model.
Accelerate has been the core component of my framework now, some of it were extracted out and merge into accelerate, e.g., #488.

There are still sometimes I find it's hard to go, for example, `load_accelerator_state` and `save_accelerator_state`. It would be better if there is an additional `get_accelerator_state`, `set_accelerator_state` so that I can wrap it inside my checkpoint.
There were a few more examples in my mind... but unfortunately I forget them now.... I'll come back later.