Update prob query section of the doc following the removal of prob macro

[sunxd3]

Ref TuringLang/DynamicPPL.jl#604

I couldn't find a way to replicate the behavior of prob macro. Particularly, given

@model function gdemo0()
    s ~ InverseGamma(2, 3)
    m ~ Normal(0, sqrt(s))
    return x ~ Normal(m, sqrt(s))
end

model1 = gdemo0(1.0)

For

model2 = model1 | (x = 2.0,)

loglikelihood(model2, (...)) will still use x=1.0 to compute the loglikelihood.

So I did some simplification, let me know the thoughts.

[yebai]

Thanks @sunxd3 -- it looks good.

Re the question above, if one wants to change the data inside a posterior model, he will have to call decondition and then condition again. This is arguably the correct behaviour because likelihood functions are not functions of data but functions of parameters.

[torfjelde]

LGTM though one minor change:)