Mixture with CustomDist with inferred logp fails

[ricardoV94]

Description

Reported in https://discourse.pymc.io/t/nonetype-error-on-custom-distribution/12176/7?u=ricardov94

import numpy as np
import pymc as pm

def dist_beta(a, b, loc, scale, size):
    return pm.Beta.dist(a, b, size=size) * scale + loc

with pm.Model() as m:

    left_beta = pm.CustomDist.dist(
        1,
        1,
        -0.5,
        2,
        dist=dist_beta,
    )

    right_beta = pm.CustomDist.dist(
        1,
        1,
        0.5,
        2,
        dist=dist_beta,
    )

    w = pm.Dirichlet('w', np.ones(2))
    resp = pm.Mixture('resp', w=w, comp_dists=[left_beta, right_beta])
    
m.logp()  # NotImplementedError: Logprob method not implemented for Elemwise{add,no_inplace}

[ricardoV94]

This happens because the construct_ir_fgraph considers the component measurable rewrites unnecessary (since these components don't have a value variable). The mixture class on the other hand expects the components to be measurable by the time its logp is called.