The Optimization Landscape of SGD Across the Feature Learning Strength