Function solve_triangular_frule 

pub fn solve_triangular_frule<T, C>(
    ctx: &mut C,
    a: &Tensor<T>,
    b: &Tensor<T>,
    tangent_a: &Tensor<T>,
    tangent_b: &Tensor<T>,
    upper: bool,
) -> AdResult<(Tensor<T>, Tensor<T>)>
where
    T: KernelLinalgScalar,
    C: TensorLinalgContextFor<T> + TensorScalarContextFor<Standard<T>> + TensorMetadataContextFor,
    C::Backend: 'static,
    C::MetadataBackend: TensorMetadataPrims<Context = C>,
    <C as TensorScalarContextFor<Standard<T>>>::ScalarBackend: 'static + TensorAnalyticPrims<Standard<T>, Context = C> + TensorMetadataCastPrims<T, Context = C>,

Forward-mode AD rule for triangular solve (JVP / pushforward).

Computes:

• x = solve_triangular(a, b, upper)
• dx = solve_triangular(a, db - proj(dA) * x, upper)

where proj(dA) keeps only the active triangular part (triu when upper=true, tril when upper=false).