Adding support for native Huggingface gradient checkpointing

[xhluca]

It would be nice to have support for Huggingface's gradient checkpointing capabilities.
With models like BERT, you can simply do:

model.gradient_checkpointing_enable()

Here's the docs on the subject.

I think this might be a good option for models that might not be supported by GradCache (or maybe both can be combined to save even more memory, though I'm not sure if they are compatible).

[MXueguang]

Hi @xhluca,
I am not very sure about this. maybe @luyug can take a look.