Add support to convert deepseek-r1 from fp8 to bf16 on CPU

[Lumosis]

Description

Converts weights from FP8 to BF16 for a HF model on CPU.

The script is mainly ported from https://github.com/deepseek-ai/DeepSeek-V3/blob/main/inference/fp8_cast_bf16.py. The original script is GPU-based.

Example cmd:

python3 -m MaxText.fp8_to_bf16 --input-fp8-hf-path <path/to/fp8/ckpt>
--output-bf16-hf-path <local/path/to/save/new/bf16/ckpt>

For deepseek-R1, the conversion process takes 1~2 hrs and the memory usage is below ~25 GB.

Tests

I compared the model files generated using my script with those generated using the original script by running diff command:

root@ranlihao-m1-ultramem-160:/mnt/disks/persist# diff models/deepseek_bf16 models/deepseek_bf16_ref/hf-r1-671b-bf16/

root@ranlihao-m1-ultramem-160:/mnt/disks/persist# 

There is no difference.

Checklist

Before submitting this PR, please make sure (put X in square brackets):

• I have performed a self-review of my code.
• I have necessary comments in my code, particularly in hard-to-understand areas.
• I have run end-to-end tests tests and provided workload links above if applicable.
• I have made or will make corresponding changes to the doc if needed.

[gagika]

a few comments

[gagika]

Thanks