Learnable Nonlinear Compression for Robust Speaker Verification