support adam bf16 state

[XiaobingSuper]

Description

This PR supports Adam bf16 state to reduce memory usage, we have tested some smaller models or largest models (LLM models), and we get a similar convergence even with this bf16 lower precision state. For the deepseek-v3 report, they also used the bf16 state to reduce training memory usage.

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Changes

Please list the changes introduced in this PR:

• Change A
• Change B

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

[XiaobingSuper]

@timmoon10 please help review it, thanks.

[timmoon10]

This will involve increasing the number of template instantiations by 16x. As an alternative approach, how about we cast the Adam state to FP32 and reuse the existing template instantiations? This is the approach we use for FP16 state. We can consider alternative approaches like JIT compilation if the memory and compute overhead become burdensome.

[timmoon10]

It's odd that the FP16 case involves per-tensor scaling while the BF16 case does not. This is not necessarily a problem with this PR, but it is a sign that the per-tensor scaling logic does not generalize and that its design should be improved.