Enabled high-performance Automatic Tensor Parallelism (auto TP) for the MoE models on multiple GPUs/HPUs

Enabled high-performance Automatic Tensor Parallelism (auto TP) for the MoE models on multiple GPUs/HPUs #1728