Mitigating Propagation Failures in Physics-informed Neural Networks using Retain-Resample-Release (R3) Sampling