So in case I have a problem I can define as an MDP, it is better to use POMDPs.jl rather than ReinforcementLearning.jl right?
Yes. I would start with POMDPs.jl for a known MDP.
Can I define an MDP using ReinforcementLearning.jl?
Yes. You can use the environment interface to implement an MDP and use the basic implementation of policy and value iteration in ReinforcementLearning.jl (see e.g. the car rental problem). This is one place where there is currently a bit of redundancy between those packages 
. I think it would be straightforward to implement the car rental problem in POMDPs.jl.