slow_conv3d: Use at::sum for grad_bias accumulation (#65758)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/65758
The same change has been made in conv2d, the proper algorithm is both
faster and gives more precision.
Test Plan: Imported from OSS
Reviewed By: albanD
Differential Revision: D31257872
Pulled By: ngimel
fbshipit-source-id: 6ff3a7a00a05b66f83d45cc820bd0c230cb8de6d
