[FEATURE]: support BF16 mixed precision training

[ver217]

Describe the feature

BFloat16 may be useful when training LLMs. We should support BF16 mixed precision training.

UPDATED: I found avx does not support load bf16 directly. BF16 for cpu adam kernel may be not implemented in the near future.

Possible roadmap:

• fused adam kernel support bf16 grad.
• low level zero support bf16. Since it may use grad accumulation, the accumulated grad should be fp32. So we have to leave a flag, like force_fp32_grad to control this.
• gemini support bf16.
• cpu adam kernel support bf16 grad.