feat: move AR fusion kernels from trtllm

[yyihuang]

Move All-reduce fusion kernels from trtllm to flashinfer

Requirements:

• allreduce_fusion_kernel_oneshot_lamport (https://github.com/NVIDIA/TensorRT-LLM/blob/main/cpp/tensorrt_llm/kernels/communicationKernels/allReduceFusionKernels.cu#L440)
• moereduce_allreduce_fusion_kernel_oneshot_lamport (https://github.com/NVIDIA/TensorRT-LLM/blob/main/cpp/tensorrt_llm/kernels/communicationKernels/moeAllReduceFusionKernels.cu#L213)
• We might move more AR kernels than these two for simplicity, and the trtllm-style assertion might be removed.

Changes:

• Add trtllm AR-fusion kernels
• Add python interface
• Add quantization utils
• (optional) Add trtllm checker/assertion - unused yet, but replaced by torch check. Please eval it in this code review.

To be discussed:

• maintain trtllm-style assertion/check? (currently with torch check)
• interface level at allreduce_fusion_kernel_XXXX or allreduce_fusion_op? (currently at allreduce_fusion_op)

next todo:

• compile
• minimize dependency
• add flashinfer logger, exception, check (maybe torch_check)
• design communication module unified interface
• unit test on python interface
• benchmark (optional?)
• unit test on C++ interface (not in plan)

[yzh119]

Closed for now because it introduce deep trtllm dependency and hard to maintain.
We will split the trtllm comm kernels into three pieces:

1. one and two-shot allreduce kernels (w/ rmsnorm fusion), feat: add trtllm all-reduce (non-MoE) #1096
2. low-precision allreduce kernels
3. moe allreduce kernels