Support SPPL functions with exogenous parameters for modeling GLMs

[fsaad]

The minimal example is a 2-mixture of linear regression, with known coefficients and a single real-covariate

func(x):
    U ~ bernoulli(.3)
    if U == 0:
        X ~ normal(2 * x + 3, .1)
    else:
        X ~ normal(-7* x + 2, 2)

Families of models with such "regression structure" include generalized linear models (GLMs). One can, for example, first learn a mixture-of-experts GLM using an approximate inference system and then render the learned model as an SPPL program. With this representation we can leverage SPPL's exact inference engine, which can solve problems that the majority of PPLs with approximate inference cannot solve.