How about using float16 with neon ?

@BUG1989 Thanks for your great work ! I have test with mtcnn which the performance just drop a litte, but speed up about 4x, amazing!
However, when I test with a CNN Model of fr(face recognition), the performance drop a lot so that it could not work properly!
Do you have plan to support neon with half data type on NCNN recently? 
Hope for your reply when you were in convenience! Thanks a lot !