Changes - hopefully improvements - on the design of the Cube environment

[alexhernandezgarcia]

Cube design

• The source state is not part of the object space any longer. Instead of being a regular state with value 0.0 for all dimensions, it has value -1 for all dimensions.
• I have removed the variable max_val and simply hard code it to 1.0. There is no clear reason to have this additional complexity.
• The actions now include an extra dimension to indicate whether the action is from or to source. This removes the need to comparison to [0.0, ...] with .isclose.
• Instead of asserting that the state is within bounds, the action is marked as invalid and sampling can continue.

Code improvement

• Since actions from source are simpler (Jacobian is one, conversion between relative/absolute is the identity, etc.), I have (hopefully) simplified the code of some methods.
• I include one attribute epsilon in the environment to control the boundaries for clamping the inputs to log_prob.
• The config distribution parameters are more intuitive:
  • Probabilities instead of logits for the Bernoulli distributions.
  • The actual alpha and beta that will go into the distribution (instead of the inputs to a sigmoid, then multiplied and shifted).

Test data set

• I have also added an epsilon in the generation of the grid and uniform data sets. The tests I have run show that including states with dimensions equal or too close to the boundaries has an impact on the logprobs-based metrics. See wandb report.

Fixes

• The variable name fixed_distribution in some config files was outdated and has been changed to the current fixed_distr_params

Other

• Before @michalkoziarski complains, I acknowledge that I implemented the method _get_beta_params_from_mean_variance() which is eventually not used. But I have chosen to leave it for potential future use. Please excuse me the sacrilege.
• I have conducted a partial analysis of the influence of the hyperparameters and added some preliminary conclusions in this wandb report.
• New common test: forward and backward trajectories must be reversible (5b7ce4e)
  • Note that this test surfaces issues in the Crystal, LatticeParameters and Tree environments and is thus skipped for these environments until they get fixed.

[michalkoziarski]

This looks good to me, I added some minor comments.