Prior sampler should use PriorContext, not DefaultContext

[torfjelde]

Example Thanasi Bakis in Slack:

using Turing

function loglik(θ, x)
    println("In loglik")
    
    return logpdf(Normal(θ, 1), x)
end

@model function test(x)
    θ ~ Normal(0, 1)

    if DynamicPPL.leafcontext(__context__) !== Turing.PriorContext()
        println(DynamicPPL.leafcontext(__context__))
        Turing.@addlogprob! sum(loglik.(θ, x))
    end
end

sample(test(randn(100) * 1000), Prior(), 100)

results in DefaultContext rather than PriorContext.

[thanasibakis]

Thanks!!

[torfjelde]

As noted by @thanasibakis , the above snippet now does evaluate the code using the DefaultContext after v0.31.4 since we're now re-evaluating the model to convert parameters from internal to user-facing representation.

I don't think this causes any bugs, as this is just a post-inference step, but it does cause additional computation overhead, which is an issue I also raised in the PR that introduced this change #2202, and is thus related to #2215.

No matter, we should maybe be running this value extraction using the PriorContext anyways, since we're only interested in hitting that part of the ~ pipeline.

@yebai