[FEAT] Improved PagedAttention FP8 (faster kvcache dequant v2)

[tjtanaa]

Description

This is a PR to merge https://github.com/ROCm/vllm/blob/shsanyal_develop_cpa_fp8 optimized attention.cu kernel into llama_fp8_12062024 branch.

CAVEAT

Currently the attention.cu kernel does not support block size of 32 and head size of 64.
The vLLM model unittests are failing as it uses small models e.g. Gemma, Llama which has head size of 64.

Performance over this Feature PR (#346) which is another implementation of faster kvcache dequant

The following is a benchmark_throughput results of Llama-3.1-70B with fp8 dynamic quantization and kv-cache-dtype of fp8_e4m3. For sequence input token length 2048 and output token length 2048:

Branch of vll-rocmfork	Req/s	Total Tokens/s	Output Tokens/s
main	0.29	1196.2	598.1
llama-fp8-12062024	0.28	1152.46	576.23
paged-attn-fp8 #346	0.47	1932.74	966.37
this PR	0.62	2537.03	1268.51