DummySampler samples invalid indices

[HenriDeh]

See JuliaReinforcementLearning/ReinforcementLearning.jl#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.

[johannes-fischer]

As I wrote in the linked issue, BatchSampler has the same behavior.