If you want to form clusters based on raw rho values instead of t-scores acress subjects, you can by using np.mean as stat_fun. However, I don’t quite see why you wouldn’t want to use the default ttest-acress-subjects in the case of RSA.
I’ve made an extensive tutorial on RSA which ends in doing cluster-based permutation tests for statistics, here: GitHub - wmvanvliet/neuroscience_tutorials: Some tutorials for students of neuroscience (look for “Representational Similarity Analysis using MNE-RSA” and “Statistics using a cluster-based permutation test”)