This adjusts our compression settings for xz compressed files, which
should reduce download sizes by around 5-15%, at the cost of ~44% longer
compression times. Given that compression happens almost universally in
CI, rather than something humans are directly waiting for, this tradeoff
feels worth it.

If we end up seeing this be too significant an increase in CI times,
moving the more aggressive compression solely to promote-release should
be relatively straightforward and give most of the benefits.

Currently, the order of the entries in `manifest.in` depends on the
iteration order of `copy_with_callback`, which in turn depends on
`WalkDir`, which explicitly says iteration order is unspecified. It's
possible to call `WalkDir::sort_by` to give an iteration order for each
directory's entries, but it felt better to accumulate the lines and then
sort them to a) make it more evident that this is happening, and b)
enable the copying to be parallelized in the future.