Improve dqn traveling salesman problem

[caleb-vicente]

List of things to be added:

• [x] Normalization of inputs
• [x] Output of the dqn should be a softmax?
• [x] Check sizes of the network. I would say that the first layer is too small
• [x] Would be interesting to add an embedding for the input
• [x] Add a parametric action space: https://docs.ray.io/en/latest/rllib/rllib-models.html#variable-length-parametric-action-spaces
• [x] Switch relu to a version of relu that can work with 0

[caleb-vicente]

I have tryied all this options using ray but it doesnt seem to be outperforming previous results. Not getting optimal routes