optimise memory usage in perturbator computation

[maramihali]

No description provided.

[lucasxia01]

could do this at some point. Should be able to multithread at each level pretty easily and also reduce memory consumption by doing this iteratively/reusing the same vector of vectors.

although not sure how impactful this would be since the perturbator now accounts for a relatively low percentage of folding

[maramihali]

I had an attempt at this while doing various optimisations around the structure trace (moving from a 2D array to a 1D array and building the coefficient tree in-place), it makes parallelisation trickier and since the perturbator is not a memory bottleneck maybe, there is no need for the optimisation.