Don't obfuscate URLError

[crusaderky]

• Allow downstream code to selectively catch URLError when a download fails.
  Notably, this enables downstream tests to use

@pytest.mark.xfail(raises=URLError, reason="download from the internet may fail", strict=False)
def test1():
    dataset = ml_datasets.<some function>()
    ...

which will allow for flakiness only in case of download failure, and trigger a test failure otherwise.

• Fix bug where HTTPError would be treated as a generic URLError.

[ngoldbaum]

I think this can't break any existing error handling, because except Exception blocks will still catch these errors.

This change makes sense to me.