On diffusion models for amortized inference: Benchmarking and improving stochastic control and sampling