Make MCTSTree threadsafe

[rejuvyesh]

Allows simulate to update the tree from multiple threads.

[zsunberg]

Nice! Interesting to start working on this. I have a few questions/concerns:

1. Will this slow down single threaded execution at all?
2. Do we need the locks when we are only reading from the tree instead of writing to it? (e.g. getting the best action node with UCB)
3. I have wanted for a long time to restructure this package so that it can be more easily used outside of the POMDPs.jl simulation loop. I started writing that out here https://github.com/zsunberg/MonteCarloTreeSearch.jl but have not tried to finish it. If we decide to invest a lot in performance (with, e.g. multithreading), we may want to consider restructuring the interface so that it is not so tied to POMDPs.
4. From everything I have read, locking the entire tree will result in really poor performance. What people have done successfully instead is not use any locks and just use atomic operations (see https://docs.julialang.org/en/v1/base/multi-threading/#Base.Threads.Atomic for info on atomic operations in Julia) and if it fails sometimes, it is just not that bad. I think this is what they did in alpha go.

[zsunberg]

Can you answer questions 1 and 2 in my comment? Questions 3 and 4 don't need answers for this to get merged, but can you confirm that you have thought about them.

[rejuvyesh]

Thanks @zsunberg!

1. In my small experiments I haven't seen any slowdowns, but should do more benchmarking first. This PR was more to get the ball rolling.
2. I ran into some errors when not adding some locks during reading, but I should reconfirm how many of them are actually necessary.

Regarding 4, I wouldn't go as far as saying really poor performance. It's still faster than single threading I think. But yeah, there definitely are ways to optimize this further. I went for the least effort thing partly because of 3, since I wasn't sure what's the long term plan for MCTS.jl.
We have even been building a MultiAgent version of this in MAPOMDPs.jl that supports coordination graphs.