No Wrong Turns: The Simple Geometry Of Neural Networks Optimization Paths