Enable custom paged attention kernel for Navi 3/4

[hyoon1]

Introduce custom paged attention kernel for Navi3/4

• Supports cases where head_size == 128 and block_size == 16.
• It does not support alibi_slopes and kv_cache_dtype == fp8.
• It supports gqa_ratio up to 16, and shows performance gains over the existing kernel when gqa_ratio is 3 or higher. Therefore, it is enabled for gqa_ratio values between 3 and 16.
• Fixed paged attention unit test to pass on Navi

Performance Gain
Script: python ./benchmarks/benchmark_throughput.py --model --trust-remote-code --dataset <ShareGPT_V3_unfiltered_cleaned_split.json> --num_prompts 1000 --max-model-len 4096 --gpu-memory-utilization 0.95

Navi 3

Models	Num Heads	GQA Ratio	Output Token/s (original)	Output Token/s (custom)	Gain
glm-4-9b-chat	32	16	991.43	1113.75	12.3%
chatglm3-6b	32	16	1442.07	1554.23	7.8%
Meta-Llama-3.1-8B-Instruct	32	4	1143.65	1221.75	6.8%
Llama-3.2-3B-Instruct	24	3	2058.97	2146.62	4.3%
Qwen1.5-7B-Chat	32	1	904.46	882.53	-2.4%

Navi 4

Models	Num Heads	GQA Ratio	Output Token/s (original)	Output Token/s (custom)	Gain
glm-4-9b-chat	32	16	1195.56	1433.13	19.9%
chatglm3-6b	32	16	1750.34	1962.21	12.1%
Meta-Llama-3.1-8B-Instruct	32	4	1405.42	1516.69	7.9%
Llama-3.2-3B-Instruct	24	3	2419.31	2561.47	5.9%
Qwen1.5-7B-Chat	32	1	765.6	761.3	-0.6%

[shajrawi]

Thank you for the PR! can you please create it against upstream instead per our contribution guidelines? https://github.com/vllm-project/vllm