io

Since commit b2ccbbb8a4b66a718298df65ac283e2056e211fa on linux uv_uptime will (in most cases) return a fractional uptime. In itself this is fine and I personal would prefer that, but apparently an extra floor was added in the windows implementation to get same behavior on all platforms (61a8d97ad9eff8a5e8b0d05d3b421062cd5427de). So this at least should be discussed.

Edit: there is an related

What should happen if deliver_cancel raises an exception? In the current implementation, the shielded cancel scope will prevent it from propagating until the process exits, which might take a while if the crash occurred before signaling the process in any way. Maybe on exception from a user-specified deliver_cancel we should call the default deliver_cancel to kill the process? Or just kill() since

PyTorch core has download function and torch.hub.download_url_to_file.
Torchaudio can use it for dataset download and does not need to maintain its own torchaudio.datasets.utils.download_url.

In addition to that, there seems to be something wrong with torchaudio.datasets.utils.download_url.
It fails to download multiple of huge files. I think we should get rid of `torchaudio.datasets.ut