[EP] Support vLLM; high-throughput and low-latency

[CalebZ9909]

Description

Please include a summary of the changes and the related issue.

Fixes # (issue)

Type of Change

• Bug fix
• New feature
• Documentation update

How Has This Been Tested?

Include any tests here.

• Unit tests
• Integration tests
• Manual testing

Checklist

• My code follows the style guidelines, e.g. format.sh.
• I have run build_and_install.sh to verify compilation.
• I have removed redundant variables and comments.
• I have updated the documentation.
• I have added tests.

[MaoZiming]

Great! This is probably not urgent (can delay), as discussed, would be good to see some perf numbers.

[CalebZ9909]

Sure, the perf numbers will be uploaded along with some local tests.

[YangZhou1997]

Let's remove unnecessary README and comments in the code that are generated by AI

[YangZhou1997]

A few notes:

• This PR fixed the long warmup time Make engine core client handshake timeout configurable  vllm-project/vllm#27444. We need to build from vLLM source, as this has not been merged in the recent release.
• DeepGemm uses cuobjdump for JIT (https://github.com/deepseek-ai/DeepGEMM/blob/9b680f428484625f4f35dc3617f134187c6bcd4a/csrc/jit/kernel_runtime.hpp#L36-L44), But on AWS cuda12.0, cuobjdump is not installed by default. So we need to sudo apt install nvidia-cuda-toolkit -y
• The default cuobjdump is installed on /bin instead of $CUDA_HOME/bin, so we need to sudo ln -s /bin/cuobjdump /usr/local/cuda-12.8/bin/cuobjdump

[MaoZiming]

I am thinking adding them to the doc under ep.

[YangZhou1997]

Sounds good to me! Let's use docs to keep consistent with the main docs folder in UCCL repo