Support twoshot kernel

[yizhang2077]

Motivation

support twoshot allreduce mainly from tensorrt llm, so custom allreduce can support larger data size, correctness has been verified, performance result in a100:

test_size = 512, world_size = 2, vllm time = 0.0130ms,custom time = 0.0119ms
test_size = 4096, world_size = 2, vllm time = 0.0113ms,custom time = 0.0136ms
test_size = 32768, world_size = 2, vllm time = 0.0121ms,custom time = 0.0122ms
test_size = 262144, world_size = 2, vllm time = 0.0176ms,custom time = 0.0170ms
test_size = 524288, world_size = 2, vllm time = 0.0223ms,custom time = 0.0228ms
test_size =1048576, world_size = 2, vllm time = 0.0325ms,custom time = 0.0326ms
test_size = 2097152, world_size = 2, vllm time = 0.0553ms,custom time = 0.0543ms

test_size = 512, world_size = 4, vllm time = 0.0135ms,custom time = 0.0119ms
test_size = 4096, world_size = 4, vllm time = 0.0121ms,custom time = 0.0140ms
test_size = 32768, world_size = 4, vllm time = 0.0135ms,custom time = 0.0137ms
test_size = 262144, world_size = 4, vllm time = 0.0239ms,custom time = 0.0206ms
test_size = 524288, world_size = 4, vllm time = 0.0313ms,custom time = 0.0306ms
test_size = 1048576, world_size = 4, vllm time = 0.0471ms,custom time = 0.0455ms

test_size = 512, world_size = 8, vllm time = 0.0140ms,custom time = 0.0137ms
test_size = 4096, world_size = 8, vllm time = 0.0142ms,custom time = 0.0165ms
test_size = 32768, world_size = 8, vllm time = 0.0186ms,custom time = 0.0189ms
test_size = 262144, world_size = 8, vllm time = 0.0300ms,custom time = 0.0287ms
test_size = 524288, world_size = 8, vllm time = 0.0354ms,custom time = 0.0348ms
test_size = 1048576, world_size = 8, vllm time = 0.0531ms,custom time = 0.0515ms

end to end performance will be analyzed in #2511

Modifications

Checklist

• Format your code according to the Contributor Guide.
• Add unit tests as outlined in the Contributor Guide.
• Update documentation as needed, including docstrings or example tutorials.