Think about parallel execution of our mapAccums and their derivatives (associative operations?)

[Mikolaj]

When the implementation of fold, scan, mapAccum and their derivatives settles down, let's think about parallel execution. A GPU backend to actually benchmark the result wouldn't hurt either. Tom provides

https://jax.readthedocs.io/en/latest/_autosummary/jax.lax.associative_scan.html

At the time of writing, everything is a special case of mapAccumL or mapAccumR, but it may very well make sense to re-introduce a standalone fold just to get a better parallelism (overriding a default method may be a good start) and then we may want to generalize the type of the fold method to any heterogeneous vectors of tensors, which by that time may get easier to use than currently (via typing, adaptors or otherwise). If if turns out scan is easier to parallelise than mapAccum (unlikely?), we may also recover the derivative expressed in terms of scan and build instead of in terms of mapAccum, at the cost of applying the derivative of the little function twice (details in old commits).