Add code needed for basic test distributions in a differentiable way (e.g two moons, roll)

[EiffL]

This is issue is to discuss which distributions we want to implement and to setup a basic package. I'll be working on it on this branch basic_infra.

[EiffL]

One thing that I noticed when adding the swiss roll distribution is that we have been using in the past with @b-remy a version that had some stochasticity, as it was based on a sampling from the underlying scipy distribution. Not great....