introduce cache array to make discrete-time state update atomic

[baggepinnen]

One step towards solving #2701

system in issue above now produces the correct answer

[baggepinnen]

@AayushSabharwal the test failure indicates that the indices used to index into p are of some magical type? Tests pass locally so I assume this is some recent change upstream?

[AayushSabharwal]

Nope this is all in MTK. ParameterIndex is an index into MTKParameters. disc_range isa Vector{<:ParameterIndex}. I didn't notice the error, but we don't have getindex(::MTKParameters, ::Vector{<:ParameterIndex}) defined, so this call is invalid. I'm not sure how it passes locally.

I guess we either need a method for the above, which just broadcasts the getindex or do it manually here

[baggepinnen]

What's the difference between these two then? The view version is used in a lot of places

cache = p[$disc_range]
disc_unknowns = view(p, $disc_range)

[AayushSabharwal]

That's interesting. I'll look into why view works, but I suspect it is because view internally just relies on getindex working for scalar indices.

[baggepinnen]

ah, I see the difference now, I'll push a fix

[AayushSabharwal]

So view won't work either. There are a bunch of branches based on whether has_index_cache(sys) && get_index_cache(sys) !== nothing (which is basically a check for if the parameter object is an array or MTKParameters). The view version is never used for MTKParameters.

[baggepinnen]

There are likely additional compiler-related problems, outlined in #2701