Drop implicit broadcasting of observations

We currently allow this:

julia> @model function f(vector, matrix)
           mean ~ Normal()
           vector ~ Normal(mean)
           matrix ~ MvNormal(fill(mean, 2), I)
           return nothing
       end

julia> f([1.0, 2.0], [1.0 2.0 3.0; 4.0 5.0 6.0])

In other words, we implicitly broadcast ~ statements when the RHS is of lesser tensor rank (scalar, vector, matrix, etc.) than the LHS. We only do this for observations, for assumptions we require the use of .~ for univariate distributions or loops for multivariate ones.

Are we sure we want to keep supporting this long-term? I find it confusing for the same reasons I described in https://github.com/TuringLang/DynamicPPL.jl/issues/825#issuecomment-2690352775, namely that it muddies the distinction between distributions of different tensor rank and goes against Julia's broadcasting conventions, such as disallowing things like [0.1, 0.2] + 1.0. Also, the univariate cases can trivially be changed to use .~ instead. The multivariate cases would require loops.

This issue is kinda the mirror image of #825. This arose originally from a question of why does tilde_observe call loglikelihood when tilde_assume calls logpdf here.

Start by comparing the observation path through tilde_observe and loglikelihood with the assumption path through tilde_assume and logpdf, then read the discussion in issues #825 and #2549. Determine whether scalar, vector, and matrix observation broadcasting should remain supported, and define the migration scope for univariate and multivariate cases.