[Feature] Add bf16 adamw

[zhtmike]

This is the optimizer used in our movie-gen & inner cogvideox project, extracted as an independent PR. It solves the convergence issue in our movie-gen model without using Mindspore AMP interface (it just takes too much memory. :( ), as @hadipash mentioned, this optimizer is very useful.

Reference:

DeepMind: Scaling Language Models: Methods, Analysis & Insights from Training Gopher section C.2

We found it best to maintain float32 parameters purely for the optimiser update. One can
partition the set of float32 parameters for optimisation updates alone along with the optimiser
state as in Rajbhandari et al. (2020). The float32 parameters are used for the update and
again cast to bfloat16 for the forward pass. This matches performance of full float32
training, improves the speed, and has only a slightly increased memory footprint compared to
bfloat16 training.

What does this PR do?

Fixes # (issue)

Adds # (feature)

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline?
• Did you make sure to update the documentation with your changes? E.g. record bug fixes or new features in What's New. Here are the
  documentation guidelines
• Did you build and run the code without any errors?
• Did you report the running environment (NPU type/MS version) and performance in the doc? (better record it for data loading, model inference, or training tasks)
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

@xxx