Broadcast `lhs .~ f(args...; kwargs...)` crashes with MethodError when mixing positional and keyword args

[docxology]

Summary

The broadcast form of the model DSL raises MethodError (no method matching tuple)
at model-construction time when the RHS supplies both positional and keyword arguments,
e.g. z .~ f(μ, σ = σ). Keyword-only and positional-only broadcast work; only the mixed
case is broken.

Root cause

src/model_macro.jl:617-628 combine_broadcast_args(args::Vector, kwargs::Vector) builds the
keyword NamedTuple from the entire broadcast-argument tuple args instead of just the
keyword values:

GraphPPL.MixedArguments((μ,), NamedTuple{(:σ,)}(args))   # args = (μ, σ) → MethodError

Repro

using GraphPPL, Distributions
import GraphPPL: @model, create_model

@model function bc_mixed(x, y, out)
    out ~ Normal(x, y)
end
@model function bc_main2()
    local μ, σ
    for i in 1:5
        μ[i] ~ Normal(0, 1); σ[i] ~ Gamma(1, 1)
    end
    z .~ bc_mixed(μ, σ = σ)     # positional μ + keyword σ → MethodError
    out ~ Normal(z[5], 1)
end
create_model(bc_main2())        # ERROR: MethodError no method matching tuple(...)

Suggested fix

Select only the keyword slice of the broadcast args for the NamedTuple in the else branch
of combine_broadcast_args (keep positional values in MixedArguments.args). Add a
regression test for the mixed broadcast case.