Add a way to override layer choices

[odow]

We need a generic way to override layer choices. This is particularly important for Gurobi to be able to opt-in to sigmoid/max constraints, and to be able to swap out ReLU formulations in a NN.

I currently have a hacky config argument in (F)lux, but that doesn't translate to GLM etc.

    historical_df = read_df("college_student_enroll-s1-1.csv")
    model_glm = GLM.glm(
        GLM.@formula(enroll ~ 0 + merit + SAT + GPA),
        historical_df,
        GLM.Bernoulli(),
    )
    application_df = read_df("college_applications6000.csv")
    n_students = size(application_df, 1)
    model = Model(Gurobi.Optimizer)
    set_silent(model)
    application_df.merit = @variable(model, 0 <= x_merit[1:n_students] <= 2.5)
    application_df.enroll = MathOptAI.add_predictor(
        model,
        model_glm,
        application_df,
        MathOptAI.Sigmoid() => Gurobi.Sigmoid(),  # How???
    )