This PR separates the logic to perform a depth-first traversal on module graph out of InnerModuleLinking/InnerModuleEvaluation. It has two advantages:

• it makes it obvious that traversal for linking and evaluation work the same way, since it's not copy&pasted code that needs to be kept in sync
• especially for evaluation, it disentangles the traversal logic from the "keeping track of top-level await consequences" logic.

Whether this is an improvement or not is subjective (it adds a layer of indirection), so I'm
curious to read the editors opinion about it.

From an editorial/formatting point of view, ModuleGraphDFS is an AO that takes a bunch of callbacks. I've declared them inline in the Link()/Evaluate() concrete methods as abstract closures, even though they don't actually close over anything. I think having them inline rather than as separate AOs makes the reading flow simpler. Ideally, my preferred approach would be something that lets us write

1. Perform ModuleGraphDFS(_module_, _initialStatus_, _pendingStatus_) with
     Action(_m_):
       1. ...
     SCCCompletion(_m_, _sccRoot_):
       1. ...
     PostErrorCleanup(_m_, _errorCompletion_):
       1. ...

I suggest reading this commit-by-commit. The first two commits just move logic around so that what would go in the same AC of the third commit is already all in the same place.

This PR preserves the bugs that would be fixed by #3613 and #3583, but by implementing this PR + those two in engine262 all the tests I have still pass.

The import defer proposal would add a new callback to ModuleGraphDFS, _getModuleDescendants_, that for linking just returns the RequestedModules while for Evaluation it uses import defer's step 11 of https://tc39.es/proposal-defer-import-eval/#sec-innermoduleevaluation (with GatherAsynchronousTransitiveDependencies).