No Trick, No Treat: Pursuits and Challenges Towards Simulation-free Training of Neural Samplers