Expose FSDP2 MixedPrecisionPolicy params

[EugenHotaj]

It would be a good user experience improvement to expose FSDP2 MixedPrecisionPolicy to be set through the config, at least for param_dtype and reduce_dtype. These are important parameters when training in low precision (e.g. bf16) and right now are only changeable by hardcoding training.fully_shard. See #2254 for why these are important parameters.

As a suggestion, we may want to hardcode reduce_dtype=torch.float32 by default. I don't think it reduces training speed at all (but we should check) and helps with convergence / stability.

[joecummings]

Agreed! Our usual approach is to limit the scope of APIs until such time as there is reasonable demand. Based on your experiments, I'd say this is very reasonable :)