Add better mechanism for recreating pods on Helm upgrade

[fwcd]

We currently use a timestamp annotation in our deployment template (similar to the suggestion in https://stackoverflow.com/a/64151965) to ensure that the helm upgrade run by GitHub Actions always recreates the pods. However, terminating a D2 pod sometimes takes up to a minute, during which both the old and the new instance may respond to user commands.

Ideally, we could find a way to either wait for the old pod to fully terminate (with the downside of having ~1m downtime per deployment), coordinate termination in a way that the terminating pod no longer responds (perhaps even by using Discord's sharding mechanism?) or restart the pod directly after running helm upgrade.