Replace `missing`-based latent selection with preallocation and conditioning

DynamicPPL uses missing in model arguments and conditions to decide whether a tilde statement is latent or observed. This couples stochastic role to mutable runtime data. Scalar missing can denote a latent aggregate draw, while arrays containing missing may instead reach loglikelihood and fail.

Deprecate and remove missing as a latent-variable marker. Models should be unconditioned by default. Structured variables can use ordinary preallocation, which current main supports, and observations should use condition:

@model function f(n, ::Type{T}=Float64) where {T<:Real}
    x = Vector{T}(undef, n)
    x ~ product_distribution(fill(Normal(), n))
end

latent_model = f(n)
observed_model = condition(latent_model, x=data)

Passing missing to a stochastic site should eventually raise a migration error.

Open issues #969, #1361, #944, and #273 show that missing-specific failures span aggregate observations, mutation, copying, and submodels. Closed issue #641 and earlier fixes #191 and #617 show that this mechanism has repeatedly required special cases. Replacing it would address their shared cause while complementing the explicit-role direction discussed in #519 and #965.

Start with the model and condition entry points, then trace how missing reaches stochastic sites, aggregate observations, mutation, copying, and submodels. Use the behavior described in issues #969, #1361, #944, and #273 to identify affected paths. Done means ordinary preallocation and condition handle observations, while passing missing produces the planned migration error.