updating upsampling bilinear2d kernel: (#21879)
Summary:
1. faster atomicAdd trick for fp16 backward kernel
2. better launch configs for backward kernel
3. removed unnecessary buffer initialization for forward kernel
Pull Request resolved: https://github.com/pytorch/pytorch/pull/21879
Differential Revision: D15898680
Pulled By: ezyang
fbshipit-source-id: 1fc81e6c078f1538d82e4f36921b630499eb504f