fix a bug in the softmax function. It executes backward comptuation incorrectly (in the condition of cpu computation and NOLOSS).
I found this bug by chance, then I spent one night confirming that something must be wrong and I spent one night finding the location of bug.
Finally, I fixed this bug in five minutes.

1. add gather function 2. add cross entropy forward computation and backward computation 3. code optimization 4. merge with xiaotong-working branch

1. add dropout function 2. add some function in tensor/core/math/unary file 3. merge with xiaotong-working branch

1. add macro to implement unary function 2. add sub and div function 3. merge code with the latest branch of xiaotong-working

1. Bug fixed; 2. Support abs, sign, convert data type(float <->int) ;3. Merge with xu and xiao branch