Implement Multi-Armed Bandit adaptation for SGMCMC algorithms

[rlouf]

Blackjax has two stochastic gradient algorithms implementations, but to use them one needs to specify hyperparameters manually, namely the batch size, step size and schedule for SGLD, plus the number of integration steps for SGHMC. @jeremiecoullon has proposed an adaptation scheme MAMBA which on top of relieving users from having to choose hyperparameter values, seems to lead to good values.

Paper: https://arxiv.org/abs/2105.13059 Code: Github