This replaces our use of Circle CI's arm64 Linux runners for both testing and releasing with the runs-on runners that Pantsbuild configures (and pays for) in AWS, via https://runs-on.com

This has a few benefits:

• single tool to learn for configuring CI
• single source of credentials (currently the release build is failing due to a token seemingly being invalid, but it's not obvious to me how to fix this)
• more atomicity, e.g. instead of the release needing to trigger a separate Circle CI build that runs in the background, we can just use the normal GitHub "wait for another job" rules

It also has downsides: we're going from free resources to paid. However, the incremental cost for the scie-pants repo should be quite small, compared to the main pants repo:

• this repo is low volume (over the last 12 months we've done about 220 CI runs total)
• things run quickly (in this PR, ~7 minutes without a cache, ~4 minutes with)

I estimated cost based on US$0.1344/hour for a on-demand t4g.xlarge 4 core instance in us-east-1 (I'm not sure this is our exact configuration). Prices in practice may be cheaper, because Runs-On attempts to use spot instances).

• per build, we're looking at US$0.0090 (with cache) to US$0.016 (without cache)
• per month, we're looking at ~20 builds, so US$0.18 to US$0.32 (but likely cheaper given this is on-demand costs)

If this starts adding up a lot, we can likely optimise. For instance: optimise spot-instance selection for lowest cost not "PCO", and use more caching (whether that's coarse-grained actions/cache or fine-grained sccache).

Fixes #422, fixes #430