Integrate PagedAttention Optimization custom kernel into vLLM

[lcskrishna]

This PR introduces the custom optimized PagedAttention on ROCm. To use the Custom PagedAttention use the following env variable VLLM_USE_ROCM_CUSTOM_PAGED_ATTN=1.

Currently this PagedAttention kernel is supported with fp16 datatype, with GQA ratio=1 to 16 and Head-size=128 and block-size=16.

Kernel was authored by : @sanyalington

cc: @shajrawi @sunway513

[shajrawi]

1. Can we have VLLM_USE_ROCM_CUSTOM_PAGED_ATTN on by default? i.e. need to be set to false to use the old one? That's what we do with Triton vs CK
2. Can you add the env variable setting to our performance guide? https://github.com/ROCm/vllm/blob/main/ROCm_performance.md

[sanyalington]

Paged attention custom reduction kernel supports a max context length of 16K beyond which it can produce incorrect results. A fix for this and further reduction kernel perf improvements are going to be available in a later drop.

[lcskrishna]

@shajrawi added into the documentation and enabled by default. The context length <= 32k support will be added into the next PR after this one.

[shajrawi]

lgtm