DeepSeek V4 FP8 training some parameters need keep bfloat16

**Describe the bug**

```
--fp8_recipe blockwise \
--fp8_format e4m3 \
--fp8_param_gather True \
```

我们这边发现，FP8训练保存下来的模型。
- compressor 的 wgate wkv 也被量化训练了。但是官方ckpt是bf16
- 另外 ape 是fp32的，开启bf16的时候这部分的参数也变成fp32。官方ckpt是fp32

参考：https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-Base/tree/main?show_file_info=model-00004-of-00046.safetensors

**Steps/Code to reproduce bug**

Please list *minimal* steps or code snippet for us to be able to reproduce the bug.

A helpful guide on on how to craft a minimal bug report http://matthewrocklin.com/blog/work/2018/02/28/minimal-bug-reports.

**Expected behavior**

A clear and concise description of what you expected to happen.

**Additional context**

Add any other context about the problem here.