Function nnls3

tests/fl/gfx/rgbw_colorimetric.cpp:733–760

Port of xy_target_rgbw_model.py `_nnls_solve` — projected-gradient NNLS for a 3x3 system. The Python version prefers scipy.optimize.nnls when available; we use the same projected-gradient fallback (500 iters @ step 0.01) since scipy isn't available in this TU. The reference falls back to this code on platforms lacking scipy, so this fidelity matches the reference's actual portable behavior.

// to this code on platforms lacking scipy, so this fidelity matches the
// reference's actual portable behavior.
inline void nnls3(const float M[3][3], const float b[3], float t_out[3], float& residual) {
    float t[3] = {0.0f, 0.0f, 0.0f};
    for (int it = 0; it < 500; ++it) {
        // r = M·t − b
        float r[3];
        for (int i = 0; i < 3; ++i) {
            r[i] = M[i][0]*t[0] + M[i][1]*t[1] + M[i][2]*t[2] - b[i];
        }
        // grad = Mᵀ·r
        float grad[3];
        for (int j = 0; j < 3; ++j) {
            grad[j] = M[0][j]*r[0] + M[1][j]*r[1] + M[2][j]*r[2];
        }
        // t ← max(t − step·grad, 0)
        const float step = 0.01f;
        for (int j = 0; j < 3; ++j) {
            float v = t[j] - step * grad[j];
            t[j] = v > 0.0f ? v : 0.0f;
        }
    }
    // residual = ‖M·t − b‖₂
    float r[3];
    for (int i = 0; i < 3; ++i) {
        r[i] = M[i][0]*t[0] + M[i][1]*t[1] + M[i][2]*t[2] - b[i];
    }
    residual = fl::sqrt(r[0]*r[0] + r[1]*r[1] + r[2]*r[2]);
    t_out[0] = t[0]; t_out[1] = t[1]; t_out[2] = t[2];
}

// Port of `_strict_project_target_xyz_to_led_hull`. When target_xy is outside
// the LED RGB triangle, projects to the nearest in-hull point by trying NNLS