Refactor the FP8 recipe implementation

**Is your feature request related to a problem? Please describe.**
Right now we support FP8 blockscale GRPO in dense & MoE, and FP8 KV quant in rollout with scales calibrated from training. The code size for these support is bloating in one file, we need to cleanup this and determine the interface for adding future support of new recipes.

**Describe the solution you'd like**
A cleaner interface of quantized recipe and a cleanup of current code.

**Describe alternatives you've considered**
A clear and concise description of any alternative solutions or features you've considered.

**Additional context**
Add any other context or screenshots about the feature request here.