This is a bit tricky as I dont know if the log likelihood function of p(P1 > P2) is easily expressed.
If you can find out the logp function for this you can write it as a DensityDist. Otherwise, you will need some likelihood-free inference approach, which we are currently working on implementing it.