Here's a draft PR in case we need it during sprints. We can always revert after the sprints if necessary.

The Cirrus workers are much faster than GitHub ones. Here's the free-threaded job (from April), with config cache available.

Left is Cirrus, 4m 16s total
Right GHA, 7m 54s total

The regular build is similar: Cirrus 4m 21s, GHA 6m 47s.

The Cirrus workers have 4 CPUs compared to GH's 3.

However, when things are very busy, like during the sprints, we're limited by the number of Cirrus workers available. I don't have a number but some runs today looked like this:

The blue lines are the normal limiting factor: Windows/FT takes about 30 minutes, and everything else normally finished before this.

But here it took about 43 and 46 minutes for the two Cirrus jobs to run. There's also a GH macos-13 in this picture, but it started almost immediately and finished after 10 mins.

And we have lots of capacity now with GH Enterprise: 50 max concurrent macOS runners:

So the speed of Cirrus doesn't matter here when we're waiting 45 mins for them to even start.