Regulate flash_attn_varlen_fp8_pertensor_func according to precision issue

[apinge]

Motivation

Modifications

Accuracy Tests

Benchmarking and Profiling

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.

[sammysun0711]

#125 refactor kv quantization to be fused in mrope rms kernel, keep this on hold if new approach can fix Qwen3-MoE accuracy issue.