Update ds script to support large block_size, safetensor and async-scheduling

[wuhuikx]

This PR has been merge into vllm main and rocm/vllm dev/perf vllm-project#27224

Then we can support block_size > 1.

[wuhuikx]

@ZhiweiYan-96 @zhuyuhua-v @gbyu-amd

[zejunchen-zejun]

LGTM

[wuhuikx]

vLLM dev/perf branch Version: 0.11.1rc2.dev0+ge9fce7brocm701
AITER dev/perf branch Version: 0.1.5.post5.dev5+gd0a40f55c

1. MI308 DS blockscale. launch_deepseekr1.sh with VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS=0

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
gsm8k	3	flexible-extract	5	exact_match	↑	0.9204	±	0.0075
		strict-match	5	exact_match	↑	0.9189	±	0.0075

2. MI308 DS blockscale. launch_deepseekr1.sh with VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS=1

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
gsm8k	3	flexible-extract	5	exact_match	↑	0.9212	±	0.0074
		strict-match	5	exact_match	↑	0.9212	±	0.0074

3. MI308 DS PTPC. launch_deepseekr1_ptpc_fp8.sh with VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS=0

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
gsm8k	3	flexible-extract	5	exact_match	↑	0.0091	±	0.0026
		strict-match	5	exact_match	↑	0.0000	±	0.0000

4. MI308 DS PTPC. launch_deepseekr1_ptpc_fp8.sh with VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS=1

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
gsm8k	3	flexible-extract	5	exact_match	↑	0.9333	±	0.0069
		strict-match	5	exact_match	↑	0.9356	±	0.0068

[wuhuikx]

@tjtanaavllm @kliuae-amd it's strange that the blockscale has accuracy regression. It should be 0.95 right?