accumulator couplers need to trigger update on dependency during invalidate()

[tclune]

Currently Accumulator actions accumulate the current value of the input during the invalidate() step. This works when the coupler is directly connected to the user output. But if there is an intervening coupler the "lazy" strategy for updating values interferes with the functioning of Accumulator.

Probably the fix is to add a method to GenericCoupler that controls whether to call update() on dependencies during invalidate(). Who knows, some other coupler might need the same.

[tclune]

There might be a sneaky way to do this without changing the general logic: if we can make the dependent coupler part of the AccumulatorAction, then we can call the update() there. Still lots of code to change, but it is at least isolated to the subclass that breaks the pattern.

[darianboggs]

Capturing Thoughts/Questions

• How to set stale?
• When to set stale?
• invalidate on a coupler may initiate update on sources
• invalidate on a coupler initiates invalidate on consumers
• update on a coupler initiates update on sources
• update on a coupler may initiate invalidate on on consumers
• Must avoid dependency loops
• Are dependency loops an issue for GridComps in general?
• Can we insist that a coupler chain be directed with regular GridComps at start and finish?

[tclune]

Unless we also have update() generate invalidate() there is 0 chance of loops. Couplers grow like a tree. (Technically a DAG because of weirdness with vertical).

Worst case scenario - invalidate() always triggers update() and we waste computations. (We very much want to avoid that though.)