Don't cache subelements

[watt]

Summary

In LayoutStorage we cache subelements, under the assumption that they will remain stable during a layout pass. This PR removes the caching, because it was a faulty assumption.

Details

In LayoutStorage we currently assume the array of LayoutSubelements for subelements will be stable during a layout pass, and cache it on the LayoutTreeNode. This was assumed safe because LayoutStorage-backed themselves elements cannot vary during a layout pass, and LazyStorage, which can, does not cache its subelement at all.

Turns out there is a case where this assumption is violated, if a LazyStorage produces a subtree that varies somewhere in a descendant node. For example, a GeometryReader that contains a stack, and varies the number of children in the stack based on space available.

Performance

This PR is broken into 2 commits:

• First, a fix that is guarded behind a LayoutMode option. This option was used to performance test this change against the current behavior as a baseline.
• The second commit removes the option and adopts the fix unconditionally.

I was expecting to incur a hit from removing this caching but benchmarking showed it to be completely negligible, and attempts to more cleverly cache actually performed worse. I'll put up a Market PR that shows this benchmark and a regression test.

Testing

There's a new GeometryReader test to validate this case, and regression test will be on the Market side.

See UI-4599.

[kyleve]

Thanks for getting this through!

[watt]

Details of the alternatives I attempted:

• Caching each LayoutSubelement by identifier. This didn't work because child.content may vary and need to be re-evaluated.
• Making LayoutSubelement capture child.content as a closure and lazily evaluate it once the first time it is needed. Had no impact on performance.