🚀 Feature

Add Conv3dInt8 with Fbgemm backend.

Motivation

Currently our quantized conv ops only support 2d conv. In order to support video understanding models, we have to add quantized version of 3d conv. In caffe2 we have the quantized 3d conv implementation with Fbgemm backend. So we can add the quantized Conv3dInt8 with Fbgemm backend here.

Pitch

Alternatives

Additional context

cc @jerryzh168 @jianyuh @dzhulgakov @raghuramank100 @VitalyFedyunin @ngimel @mruberry