Faster versions of fvec_op_ny_Dx for AVX2 (facebookresearch#2811)

Summary:
Pull Request resolved: facebookresearch#2811

Use transpose_AxB kernels to speedup computations.

Differential Revision: https://internalfb.com/D44726814

fbshipit-source-id: 58396963215e758e42f518ad4e556e9dbf8ff7ec

Author: alexanderguzhva