The not working should be independent on pymc3 version and depend on ArviZ version. The error is triggered by the conversion to InferenceData to calculate ess and rhat at the end of sampling, thus as you mentioned, it does work without convergence checks.
As I mentioned in the github issue with the same question, I’d recommend using pm.Potential instead of densitydist for this.
We are trying to understand the root of the issue, having parameters passed as observed argument strikes me as very odd.