Simplify deferred function, class, and accessor checking

[ahejlsberg]

We defer checking of function, class, and accessor bodies that are contained in expressions in order to avoid type checking circularities. For example:

const foo = function () {
    const s = foo();
    return "hello";
}

Here, performing a full type check of the body of the function expression whilst in the process of determining the type of foo would cause foo to be given type any because of the recursive reference. Delaying the type check of the body ensures foo has been assigned a type.

We previously performed an AST walk to find the deferred function, class, and accessor bodies. The walker was a rather brittle piece of logic with lots of cases to reason about (and possibly forget). This PR replaces the logic with a simple list of deferred nodes to be checked later. This removes a bunch of code and improves checker performance by 5% or so (with the Monaco project) due to the elimination of the extra AST walk.

The baseline changes are caused by the fact that we now never check the expression in an out-of-place return statement (as intended). We previously sometimes checked it (specifically in cases when the return statement contained a deferred function, class, or accessor body).

[DanielRosenwasser]

deferredCheckNodes and leave a decent comment about why these need to be deferred until the end.

[DanielRosenwasser]

Are you ever going to not have deferred nodes for most purposes? Seems like you'd have more overhead on a large project from checking over and over whether to defer than to create a one-time array allocation.

[ahejlsberg]

The deferredNodes array is used only when doing a full type check. When called from the language service (i.e. when "helicoptering in") we don't use it.

[sandersn]

👍

[DanielRosenwasser]

👍