fix: remove descriptors squeeze() to fix 1D-BD tasks

[manon-but-yes]

This PR aims to fix the container addition for 1D-BD tasks. With the current container implementation, 1D-BD tasks such as the Hopper-uni task for example (hopper_uni) returns the following error:

This is due to the squeeze of the descriptors done in the add() method of the repertoire. These squeeze are necessary for the fitnesses and indices as these arrays are explicitly expand_dims at the beginning of the add() method. However it is not needed for the descriptors that are not modified in the add() method.
Thus as a fix I propose to remove the squeeze for the descriptors.

This change should not impact the results, but just in case I attached the experimental comparison on GPU with the original container implementation and the new proposed container implementation on the ant_omni and walker2d_uni tasks. I ran 5 replications, the difference between runs is due to the non-deterministic nature of computation on GPU. I also added the results for the hopper_uni after the fix.

[limbryan]

This issue should be present in MOME repertoire as well. To be confirmed with @felixchalumeau. The corresponding line can be found here: https://github.com/adaptive-intelligent-robotics/QDax/blob/develop/qdax/core/containers/mome_repertoire.py#L293

[felixchalumeau]

Indeed! I fixed it in MOME as well by precising the axis that must be squeezed. Also, I add parametrized tests for MAP-Elites and MOME to be sure that 1 and 2 dimensions BDs are tested.