Reenable PriorContext for Optimization

[alyst]

This PR reenables the support for PriorContext in optimization, which was disabled by #2022.
While it is not directly supported by the current optim_problem() interface (only MLE/MAP are supported, although adding MaxPrior is possible), I think it is useful to support PriorContext for Turing.OptimLogDensity().
In my code I use it for multi-objective Pareto optimization (prior + llh).

[devmotion]

This PR reenables the support for PriorContext in optimization, which was disabled by #2022.

AFAICT support for PriorContext has never existed and has never been tested. The check in the PR was only introduced to communicate this more clearly and provide a better error message.

[alyst]

AFAICT support for PriorContext has never existed and has never been tested.

I see. It was working before though. Do you think it makes sense to support it for some of the user use cases?

[torfjelde]

IMO if we 're allowing both DefaultContext and LikelihoodContext, I feel like allowing PriorContext also seems reasonable (i.e. all the leaf-contexts).

[mhauru]

I had one minor style change to suggest.

I also got some extra peace of mind by, in the test, checking that

            true_prior_logpdf = -Distributions.logpdf(Uniform(0, 2), a[1])
            @test Turing.OptimLogDensity(m1, prictx)(a) ≈ true_prior_logpdf

Could incorporate a more explicit check like that.

Both of these are optional, I'd be okay with the code as is. I'll leave approving for @torfjelde since there might be context (pun unintended) to this that I lack.

[mhauru]

Suggested change

	DynamicPPL.tilde_observe(::OptimizationContext{<:PriorContext}, right, left, vi) =
	(0, vi)
	DynamicPPL.dot_tilde_observe(::OptimizationContext{<:PriorContext}, right, left, vi) =
	(0, vi)
	DynamicPPL.tilde_observe(ctx::OptimizationContext{<:PriorContext}, args...) =
	DynamicPPL.tilde_observe(ctx.context, args...)
	DynamicPPL.dot_tilde_observe(ctx::OptimizationContext{<:PriorContext}, args...) =
	DynamicPPL.dot_tilde_observe(ctx.context, args...)

I think falling back on the PriorContext implementation is a bit neater. Result is the same.