linear support deepgemm

[sleepcoo]

Motivation

Use deepgemm linear gemm, dependent on #4165

Test case

#luanch server
python3 -m sglang.launch_server --model deepseek-ai/DeepSeek-V3 --tp 8 --trust-remote-code 

MMLU Average accuracy: 0.878

performence test

size	origin	deepgemm linear
Output token throughput (tok/s)	1244.04	1360.05
Total token throughput (tok/s)	1818.26	1987.82

[zhyncs]

wait for #4255

[zhyncs]

This is a duplicate. Please delete line 726 of the code.

[merrymercy]

I tried to release a new version of sgl kernel to unblock this. But I hit some errors 1a5023e. Will need @zhyncs 's help.

[zhyncs]

v0.0.4.post1 is ready

[zhyncs]

• ut can't pass on H200
• torch compile not compatible python3 -m sglang.launch_server --model deepseek-ai/DeepSeek-R1 --tp 8 --trust-remote-code --enable-torch-compile --torch-compile-max-bs 1

[zhyncs]

ref #3232

[sleepcoo]

• ut can't pass on H200
• torch compile not compatible python3 -m sglang.launch_server --model deepseek-ai/DeepSeek-R1 --tp 8 --trust-remote-code --enable-torch-compile --torch-compile-max-bs 1

wait this deepseek-ai/DeepGEMM#55

[zhyncs]

There are still many issues to be fixed in this PR, such as the condition of arch older than hopper, oom of ut, etc., but these are all minor issues that can be resolved in follow-up.