DiffSim2Real: Deploying Quadrupedal Locomotion Policies Purely Trained in Differentiable Simulation