Exploiting Active Subspaces and Global Sensitivities in Model Predictive Control