See https://github.com/JuliaReinforcementLearning/ReinforcementLearning.jl/issues/1068 . We can either fix DummySampler, or make a new full trajectory sampler. In the former case, I think it should be renamed because it's not really dummy.