[request] new method to save the fp32 model

[stas00]

Currently checkpoint saves the original fp32 weights in zero_pp_rank_* files, which makes it inaccessible to users who may need the final result to be used elsewhere w/o Deepspeed (e.g. inference or a different system).

So we need deepspeed.save_model() which will reconstruct the master weights and save them as model.bin.

For context for this need please see: #797

Thank you.

[stas00]

@ShadenSmith shared elsewhere that

each fp32 param group is stored in a flattened buffer in the fp16 optimizer here and needs to be exposed as a method. The key is to unflatten each buffer into an fp32 copy of the model params (this would be looped over each param group)

[stas00]

Any chances this new method to save the fp32 model can be added?

Users are reporting struggles when they customize their saving/loading of the checkpoints
huggingface/transformers#10789

I think there is also a problem to switch from zero2 checkpoint to zero3, correct? Which means that we really need to salvage the fp32 model weights from the zero2 checkpoint before it can be loaded in zero3.

Thank you!

@ShadenSmith, @jeffra

[stas00]

PR with solution: #892