use fbgemm's 3d group conv fast path (#29085)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/29085
Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/159
Change DNNLOWP operators to use fbgemm's new 3D groupwise convolution (D18192339)
This diff also fixes an issue when column offsets are fused into bias.
In this case, we construct ReQuantizeOutput with col_offsets == 0 and A_zero_point == 0 even if real A_zero_point is 0.
In fbgemmGroupwiseConv, when we call dispatchOutputProcessing, we shouldn't pass the original A_zero_point .
Test Plan: https://github.com/pytorch/pytorch/pull/29134
Reviewed By: dskhudia
Differential Revision: D18282373
fbshipit-source-id: 993d584e7fa8e07c74597304c0fd9386f7ed0e41