MAINT: speed up istft by using col2im (the original python code used … (#42826)
Summary:
Fixes https://github.com/pytorch/pytorch/issues/42213
The [original python code](https://github.com/pytorch/audio/blob/v0.5.0/torchaudio/functional.py#L178) from `torchaudio` was converted to a native function, but used `eye` to allocate a Tensor and was much slower.
Using `at::col2im` (which is the equivalent of `torch.nn.functional.fold`) solved the slowdown.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/42826
Reviewed By: smessmer
Differential Revision: D23043673
Pulled By: mthrok
fbshipit-source-id: 3f5d0779a87379b002340ea19c9ae5042a43e94e
