Add some helper functions for retrieving the dnf value of `${basearch}`.

Note there are many places over the codebase where we use `uname -m` or
`arch`, but we semantically mean `${basearch}`. I didn't bother
replacing them because (1) for the architectures we care about, it
actually makes no difference (if e.g. one day we want to support 32-bit
ARM or x86, then it will) and (2) given the above, the refactor risk
didn't seem worth it.

Mark the OSTree commit with the actual `${basearch}` we used to assemble
it. The major user for this will be Zincati. Putting it in the commit
metadata means this will then show up in e.g. `rpm-ostree status
--json`, which Zincati already executes today.

We could also do this directly in rpm-ostree. Though really, after
digging into this quite a bit, I lean towards making this a CoreOS-ism
for the time being. There aren't many software out there that care about
the specific `basearch` string; it's something that mostly only matters
to yum/dnf/rpm-ostree (hence why it's not really exposed in any easy
way).

As one can guess, any portable software that needs arch-specific
behaviours would use e.g. compiler flags or `uname(2)`, rather than any
kind of distro-specific concept.

I also didn't make this part of `/etc/os-release` because if Fedora does
end up shipping it there as well eventually, we'll have to carry *two*
variants of the same tag. OTOH, keeping it in the commit metadata forces
the few classes of software that care about this to be specific to
rpm-ostree systems, over which we likely have more control.

The default loader is not considered safe. Use `safe_load`, which
disables some of the more funky YAML features we don't need.

This is a common pattern to signal that anything that subclasses us is
expected to implement these methods.

This is useful for hacking on the codebase, but also in a devel pipeline
context.

Otherwise, an ill-timed `SIGINT` can interrupt `createrepo_c` while it's
creating the yumrepo and a follow-up invocation will barf on `.repodata`
already existing.

Minor follow-ups from the OSTree-in-builddir work.

- prune_builds: we don't need to remember the `ostree_timestamp` anymore
  when collecting builds.
- prune_builds: drop the `subprocess` module, since we no longer need to
  call out to the `ostree` CLI either.
- cmd-clean: drop the reference to OSTree repo objects

Use `builds_dir` instead of `build_dir` to refer to `builds/`, and
`build_dir` instead of `build_root` to refer to `builds/$id/`. This
matches the convention used in the rest of the codebase.