Dask losing workers/nthreads over time

**Describe the issue**:
I have a simple 104-worker (=nthreads) cluster. Over the course of a couple of days the cluster is now running at 85 workers. Restarting the nurse on each node manually works, but I'd like to understand if there is a more robust solution of maintaining the cluster integrity and if this is a bug or intended behaviour.

**Minimal Complete Verifiable Example**:
I don't know if that's the issue but we're running a memory-overclocked cluster tuned to raw performance over stability, and some workers are allowed to eventually segfault with about a few segfaults per day. Not easy to reproduce.

**Anything else we need to know?**:

**Environment**:

- Dask version:dask, version 2023.4.1
- Python version: 3.11.0rc2
- Operating System: Ubuntu 22.04
- Install method (conda, pip, source): pip