Feature Implement QLearner

[usaito]

New Features

• implement QLearner in the policy module, which first estimates the expected reward function (or the q function) from the logged bandit data, and then use that estimator to make new decisions on the test data (such as via argmax or softmax)
• See example 1 of in Appendix A of https://arxiv.org/pdf/2002.08536v2.pdf for the definition of QLearner implemented here
• add some tests corresponds to the inputs and performance of the QLearner class in tests/policy
• implement "Gumbel softmax trick" (see Section 4.1 of https://arxiv.org/pdf/2105.00855.pdf) in IPWLearner/QLearner/NNPolicyLearner (of obp.policy) to efficiently sample a ranking of actions from the plucket luce ranking distribution

Minor Fix

• unify some test functions to delete repetitive training of a same policy in tests/policy/test_offline_learner_performance.py to reduce the time needed to run the tests
• fix some numpy/pandas/sklearn warnings (now the tests will run without any warnings)

[fullflu]

nits (no warning happened)

I'll add a review of QLearner in the next comment.

[usaito]

@fullflu Thanks!

[fullflu]

Review of QLearner

[usaito]

@fullflu Thanks again for your thoughtful comments! I'll address some recent minor comments in the following ongoing PR.
#145