* blockReduce opt

* launch threads align to warpSize

* reduce unnecessary shared memory for broadcast reduced value

* vectorize SoftmaxKernelWithEltadd

* add fp16 constrain

* test=develop