Destroy preallocated views that the pull model never mounted (#58890)

[bartlomiejbloniarz]

Summary:
With enableMountingCoordinatorPullModelAndroid, the UI thread pulls mount transactions instead of the JS thread pushing every commit. When one commit mounts a subtree and the next commit unmounts it before the UI thread pulls, the pulled transaction contains neither the Create nor the Delete for it. The views preallocated for that subtree stay in SurfaceMountingManager, and their tags stay in FabricMountingManager's allocated view registry. The cleanup that destroys never-mounted preallocated views does not run either, because ShadowNodeFamily already marked the family as mounted at commit time. In a test app that repeatedly mounts and unmounts 1200 views, nearly all of them leaked as Java views and registry entries.

This PR changes how preallocated views are cleaned up when the pull model is enabled:

• The allocated view registry records whether each view is only preallocated, or was created by a mount (a Create or the surface root).
• ShadowNodeFamily gets onFamilyDestroyed, whose callback runs for every destroyed family, and a hasBeenMounted() getter. onUnmountedFamilyDestroyed is deprecated. It keeps its behaviour by wrapping onFamilyDestroyed and skipping mounted families.
• destroyUnmountedShadowNode destroys only views that are still preallocated-only. A view created by a mount is left to its Delete, which can be pulled after the family is destroyed.
• Each queued preallocation holds a weak reference to its family. The drain skips views whose family is already destroyed, so it does not preallocate a view that nothing would clean up.

FabricMountingManager::preallocateShadowView is removed, because the drain now registers and preallocates its views directly and no production code called it.

Changelog:

[ANDROID] [FIXED] - Fix preallocated views leaking when enableMountingCoordinatorPullModelAndroid is enabled

Test Plan:

• ShadowNodeFamilyTest: onFamilyDestroyed runs for both mounted and unmounted families and reports hasBeenMounted() correctly, and the deprecated onUnmountedFamilyDestroyed still skips mounted families. There is no open-source target for these gtests, so I built them with the NDK and ran them on an Android emulator.

• FabricMountingManagerInstrumentationTest has new cases for the following. It has no Gradle target in the open-source repository, so I ran it on an Android emulator through a local androidTest source set that is not part of this PR.

  • the drain skipping a view whose family was destroyed (pull model);
  • a committed family destroying its preallocated view (pull model);
  • a committed family keeping a view that a mount created (pull model).

  With the native changes reverted, the cases for the destroyed committed family and the drain skip fail.

• Release build of a test app on an Android emulator. The app mounts and unmounts a subtree of 1200 views in a loop, then idles. View counts at idle:

  Configuration	Java views before → after	Registry entries before → after
  Pull model	1197 → 3	1197 → 3
  Pull model + Props 2.0	1203 → 3	1203 → 3

  There were no crashes or new soft exceptions.

• yarn cxx-api-validate, yarn format-check-cpp and yarn format-check-kotlin pass.

Differential Revision: D123648108

Pulled By: bartlomiejbloniarz

[meta-codesync]

@bartlomiejbloniarz has imported this pull request. If you are a Meta employee, you can view this in D123648108.

[meta-codesync]

@bartlomiejbloniarz has exported this pull request. If you are a Meta employee, you can view the originating Diff in D123648108.