Faster versions of fvec_op_ny_Dx for AVX2 (facebookresearch#2811)

Summary:
Pull Request resolved: facebookresearch#2811

Use transpose_AxB kernels to speedup computations.

Reviewed By: mdouze

Differential Revision: D44726814

fbshipit-source-id: a1dd3e289f4ed564a5bece699bee0af88c9925b0

Author: Alexandr Guzhva