[Kernel][LoRA]Punica prefill kernels fusion

[jeejeelee]

Summary

This PR has accomplished the following tasks:

• Performed horizontal fusion on the shrink and expand kernel of sgmv, enabling support for multiple (1-3) subLoRAs.
  This optimization reduces Triton kernel calls in MergedColumnParallelLinearWithLoRA and MergedQKVParallelLinearWithLora, lowering TTFT, while also removing the unused sgmv_expand_slice kernel.
• Made minor adjustments to the configuration of the fused expand kernel, resulting in slight performance improvements.
• Updated the corresponding unit tests for the fused kernels

Performance comparison

Models expected to benefit from these changes

• serve script

vllm serve meta-llama/Llama-2-7b-hf \ 
--gpu-memory-utilization 0.90 \ 
--served-model-name llama2-7b \ 
--enable-lora --max-loras 3 --max-cpu-loras 15 --max-lora-rank 8 \
--lora-modules lora=yard1/llama-2-7b-sql-lora-test \

Models that should not be affected

• serve script

vllm serve THUDM/chatglm3-6b \ 
--gpu-memory-utilization 0.90 \ 
--served-model-name chatglm3 \ 
--enable-lora --max-loras 3 --max-cpu-loras 15 --max-lora-rank 64 \
--lora-modules lora=jeeejeee/chatglm3-text2sql-spider \
--trust-remote-code

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.
Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can do one of these:

• Add ready label to the PR
• Enable auto-merge.

🚀

[varun-sundar-rabindranath]

LGTM !

[mgoin]

Nice work, no additional comments from me!

[Isotr0py]

LGTM!

[jeejeelee]

Add minicpmv lora tests to distributed testing to avoid OOM, and remove samples with unstable generation results in TP testing (which have unstable generation results even in the current main branch)