* add exp and reciprocal for activation
* add conv2d transpose op
* add fill constant op
* add im2sequence op
* add interpolate op (including nearest and bilinear)
* add lrn op
* add split op
* add sum op
* add topk op
* add gru(int31) op
* fix bug in elementwise arithmetic op
* fix bug in conv2d op
* fix bug in dropout op
* fix bug in cast op, test=develop test=xpu