Revert D28142447: Improve BatchNorm1d performance (CUDA)
Test Plan: revert-hammer
Differential Revision:
D28142447 (https://github.com/pytorch/pytorch/commit/b2936ad8fa7c6040218e3b6f74c019fcb0675b18)
Original commit changeset: c70109780e20
fbshipit-source-id: e93f6d00d644697b106f5ea8ab79872f353b51c6