Deep Learning for Random Objects in Metric Spaces: Beyond Manifold Assumptions