Migrate conv3d from TH to ATen (CPU) (#29007)
Summary:
This is a port of the VolumetricConvolutionMM TH (CPU) implementation to ATen as `slow_conv3d`.
- [x] unfolded3d_copy & unfolded3d_acc
- [x] forward
- [x] backward
- [x] basic sanity cross check with 1.3 impl
- [ ] systematic testing
- [ ] performance comparison & optimization
Script used for performance testing: [benchmark_conv3d.py](https://gist.github.com/andreaskoepf/8865eea4bb05220f78fc6d9d408c49fc)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/29007
Differential Revision: D18299298
Pulled By: ezyang
fbshipit-source-id: 97d53e8c976a09aecbc6f05dd8e982cc58cdf6d8