onnxruntime

Optimize ReduceMean/ReduceSum when all reduce axises located at the tail of the
input tensor's dims by do not make extra copy. And use openmp to parallel the
reduce on results.