Add truncated normal distribution

[femtomc]

I'll take a swing at this today. Looks cool.

[ztangent]

@femtomc I wonder if it might be worth making a generalized truncated operation, similar to: https://juliastats.org/Distributions.jl/stable/truncate/

[ztangent]

We could do a fallback implementation like:

function _logpdf(d::Truncated, x::T) where {T<:Real}
    if d.lower <= x <= d.upper
        logpdf(d.untruncated, x) - d.logtp
    else
        TF = float(T)
        -TF(Inf)
    end
end

and specialized implementations for common truncated distributions like the truncated normal.

[femtomc]

I was discussing this a bit with Alex -- will comment in a bit.

[femtomc]

I agree with your comment. Basically, start with truncated normal. As we build more, you can specialize and have a function truncate which mimics the Distributions.jlmethod truncated.

[ztangent]

I was actually thinking the other way round! But I realized on further thought that a fallback implementation won't be as easy, just because we don't equip Gen distributions with CDFs like Distributions.jl does, and that means we can't calculate d.logtp for every distribution easily.

[marcoct]

See https://github.com/probcomp/Gen.jl/issues/362