fix: resolve GH_REGISTRY to IP in before_script

We're seeing a lot more unknown blob errors on pushing to ghcr.io (and only there; the missing blobs are missing on the server). One possible explanation is load-balancing and geo-location causing ghcr.io to resolve to different domains for multiple layer pushes in a container push, so some layers end up getting pushed to a server that doesn't have the underlying layers (yet). Normally, DNS caching should make sure that ghcr.io resolves to the same destination for subsequent calls. So, a hypothesis is that something changed in the DNS caching on eicweb (or DNS server we pull from), or caching never worked and GitHub added some servers that we now start picking up with some regularity.

This MR does our own version of DNS caching :-)