cdist performance improvement for euclidean distance (#25799)
Summary:
jacobrgardner https://github.com/pytorch/pytorch/issues/15253#issuecomment-491467128 preposed a way to speedup euclidean distance calculation. This PR is implementation of this solution for normal and batch version.
Also simonepri provided performance metrics https://github.com/pytorch/pytorch/issues/15253#issuecomment-502363581
Current implementation has speedup comparing to jacobrgardner approach
Pull Request resolved: https://github.com/pytorch/pytorch/pull/25799
Differential Revision: D17964982
Pulled By: ifedan
fbshipit-source-id: bf7bd0dbfca51fd39e667da55139347480f30a2f
