Decision on ACE Arguments

[cortner]

There are a few (related) decisions to be made:

• ACEConfig vs just anything iterable
• do we distinguish site state from (i,j) pair state? The distinction is very natural in the context of interatomic potentials but not so natural in other contexts.
• If we do go for 2 arguments, then it feel much more natural to choose (Xi, Xj) rather than (Xi, Xij). But is this a restriction of generality?

To make an informed decision on this, I think we need to benchmark the cost of converting (Xi, Xj) to a single Xij state that combines information from both.

[cortner]

Initial tests in notes/arg_benchmark.jl suggest that - if the allocation in the State could be removed, cf #89 - then the cost of converting from a structure to an environment, possibly duplicating lots of data would be negligible, on the order 50 to 100 ns for a fairly complex state variable; so not even getting close to the us range. (even without fixing the allocation it would likely be ok in the 200-500 ns range)

[cortner]

Since state allocation can very likely be fixed, I will close this now.