Towards Adversarially Robust and Domain Generalizable Stereo Matching by Rethinking DNN Feature Backbones