For all the children classes of BasePerturbation, the randomness only appears at the .predict() step. So, for clarity, I suggest we only set/check random_state starting from that stage.

In addition, it would be desirable that the behavior of the fitted class is deterministic: calling twice .importance() on the same data gives the same result.

I suggest that we expose the random_state at the importance/predict methods, but not before.

For compatibility with CFI, this requires modifying the ConditionalSampler, which currently takes the random_state at init but only uses it at .sample() we could thus make it an argument of .sample(random_state=None)

Originally posted by @jpaillard in #344 (comment)