* Fix softmax cuda bug

* Refine multihead log and softmax logic

* Align block to 32

* Improve argsort performance.

- Give 200000 data to compute argsort on v100,
can speed up ~190x
before opt cost: 0.53s
after opt cost:0.0027s

- Add fp16 support

* Refine error message
* Refine code
* Add descending sort

test=develop
Signed-off-by: Nzhaoyuchen <zhaoyuchen01@baidu.com>

The op should handle k=1024
Fix seq_len < warpsize error.

test=develop
Signed-off-by: Nzhaoyuchen <zhaoyuchen01@baidu.com>

test=release/1.6

* Add multihead op for ernie opt

* Refine softmax

* Refine kernel.

* Refine cuda kernel

* Refine cuda version

* Refine cmake