Update fp8 paged attention for MI308

[amd-xiaoyu12]

Please direct your PRs to the upstream vllm (https://github.com/vllm-project/vllm.git)

Accepting PRs into the ROCm fork (https://github.com/ROCm/vllm) will require a clear previously communicated exception

Summary:
Support full fp8 MFMA with wrap level dynamic query quantization to improve fp8 performance on MI308, which can also benefits other MI300x accelerator or latest hardware.

• Performance

• Unit test - attention output

* Lm-eval-harness ppl test