As @cluhmann I suspect the model has not converged somehow and there are some draws whose value is infinity (represented in np.inf). This is not considered nor supported in the KDE function that errors out when trying to generate the KDE with infinite bandwidth.
You should be able to confirm there are samples with infinity value using
np.isfinite(cov_..._idata.posterior).map(np.all)
Or there might also be a warning during or after sampling as @cluhmann suggests, not sure about that.