audit of core deferral code

[jbardin]

Stepping through the core deferral handling uncovered some more loose ends and repetition. Cleanup the code some more, and try to make it more understandable.

• Fix action expansion deferral recording. Action expansion was being recorded as action deferral, instead of
  action expansion deferral. We don't actually have "action deferrals" right now though, so remove that too while we're at it.
• Add missing calls to ReportModuleExpansionDeferred which were never wired into module expansion
• Remove dead code in GetDeferredPartialExpandedResource
• Make "deferrals off" behavior of Deferred structure consistent. The Report* methods still record even when deferrals are off, so the duplicate-report panics still catch bugs outside stacks
• Create a shared helper putUnique for the "insert once, panic on duplicate" logic in Deferring
• LoadPlannedDeferrals decides between a full and a partial deferral from the wildcard instance key in the address. It no longer relies on DeferredReasonInstanceCountUnknown.
• PartialExpandedResources returns the partial expansions for a resource, so resource nodes can access them during apply.
• decodeDeferredResources is the inverse of the plan's deferredResources. A missing schema or a decode failure is now a normal error diagnostic that stops apply before the walk begins, instead of an error inside a graph node.
• checkForPartialExpansion asks ctx.Deferrals() for partial expansions instead of having them pushed in by the transformer. We can remove the special transformer and extra deferral-related nodes from the graph entirely.
• unify where NodeAbstractResourceInstance defers the instance from multiple call sites.
• Fix where deferred destroys could lose track of their actions invocations.
• Fix panic from "checking whether X should be deferred when it was already deferred". Now we also check for the opposite error, that a resource can't be deferred during apply if it was not planned as such.
• Further streamline the managed instance planning process to make it easier to follow, and coalesce duplicate calls.

[jbardin]

I hate unnecessary else blocks, especially in long sections where you can't easily see which condition's else you're under, or when the primary path is the else. This makes the refresh-only path the exception, and the normal planing path the main line.

[austinvalle]

Very nice! 🧹