Add FP8 support

[MatthewBonanni]

Add FP8 support via patching in this unmerged PR from upstream: #54

[MicroZHY]

Hi @MatthewBonanni,
Thank you for porting the FP8 feature over!
Could you please summarize the main incremental changes between your branch #4 and the original upstream PR deepseek-ai#54 ?
In other words, what did you add, drop, or modify on top of the upstream patch so that it fits the vLLM fork?
Really appreciate your time!
Best regards

[MatthewBonanni]

@MicroZHY Glad to contribute! The code from the upstream PR is largely unchanged; I simply copied the relevant kernel into a new kernels_fp8 folder and updated mha_fwd_kvcache_mla to route to the appropriate kernel based on the dtype. The FP8 kernel also takes an Flash_fwd_mla_params_fp8 object, which is a child of Flash_fwd_mla_params, simply adding the descales and h_h_k_ratio, a parameter no longer used by the main kernel.

[LucasWilkinson]

LGTM! thanks