* add new api/op kthvalue

* kthvalue cuda kernel to cub sorting

* fix example code error

* throw errors instead of LOG in cuda sort

* throw errors by Paddle_ENFORCE