Issue Description:

This appears to have broken yesterday/today (diagnostic is that "property weight is not differentiable"), apparently because Differentiable.CotangentVector no longer has @FieldwiseProductSpace:

struct DenseLayer: Differentiable {
    var weight: Tensor<Float>
    var bias: Tensor<Float>

    func applied(to input: Tensor<Float>) -> Tensor<Float> {
        return matmul(input, weight) + bias
    }
}

let model = DenseLayer(weight: Tensor<Float>(zeros: [5, 5]),
                       bias: Tensor<Float>(zeros: [5]))
let x = Tensor<Float>(zeros: [4, 5])
let grad = model.gradient { m in m.applied(to: x) }

As a note, if @differentiable(wrt: (self)) is added to the properties, the snippet compiles but gets into an infinite loop at runtime between a synthetic primal and a synthetic VJP for DenseLayer.weight.