Clone tensors to avoid torch.save bloat

[tjruwase]

Fixes #3303

[tjruwase]

TODOs:

1. Docs
2. Unit tests (?)

[tjruwase]

@stas00, please see docs https://deepspeed.readthedocs.io/en/rtd-staging/model-checkpointing.html#avoiding-zero-checkpoint-bloat

[stas00]

Looking at the rendering - the source formatting appears to be borked. It has :param: and the last section doesn't show up.

And the doc is hard to read as it refers to input, let me try to make a better suggestion

[tjruwase]

@stas00, thanks for the feedback. I have applied your suggestions. Please take another look.

[stas00]

Looking good now!