Record download of completed partial requirements

[pfmoore]

Does not fix #11847 (it's a partial fix, but see that issue for further discussion).

When a requirement's metadata is downloaded from a PEP 658 metadata file, the requirement is marked as needing more preparation (because we didn't download the distribution). When we come to do that preparation, we failed to record the fact that the file was now downloaded in the _downloaded dictionary, causing the file to be downloaded a second time.

[pfmoore]

The test failures seem to be related to metadata files for source distributions. That's valid, but not a situation that I'd considered, as PyPI doesn't serve metadata for sdists yet.

I wonder if we're failing to set some crucial attribute on the req that's not used for wheels? At this point I'm out of time to investigate this further. I'll get back to it when I can, but if anyone else has any ideas, I'd love to hear them.

[pfmoore]

OK. The tests are passing, this removes the "release blocker" status of #11847 (in my view as RM), and it's as much of a fix as I'm able to commit to producing. So I'm going to merge this unless someone wants to object (and commit to working on a better fix for #11847 between now and the release, which will likely be mid-July). I'll leave it for a few hours for people to comment, but merge later today. Reverts are cheap, so if that's not enough time for people we can still revert and pass the problem onto someone else after I merge.