CI improvements - no scrambling (demultiplex) and core dumps

General issue

Currently the CI output is multiplexed by different dragonfly (master/replica/cluster etc) processes. Each dragonfly instance also multiplexes its output from multiple shards. All of this end up in a single output stream (the ci).

This unfortunately causes cluttered ci output where stack traces are scrambled, log messages and output is also scrambled or eaten out completely.

It's common that we can't know what happened just by looking at the CI output and we need to manually download the artifacts/logs.

This is a) cumbersome for us who want to quickly look logs and b) an agent would also need to pull those logs and inspect them manually.

Furthermore, we don't upload core dumps.

Tasks

[ ] demultiplex process output by redirecting to a file and on failure serialize and print what was captured. This will give a consistent word view instead of a scrambled mess. [ ] upload core dumps.

I prototyped both while working on one shot flushes. I will follow up with separate PR's and I opened this issue for tracking.