Pull model: mount the commits of starting and stopping surfaces (#59030)

[bartlomiejbloniarz]

Summary:
With enableMountingCoordinatorPullModelAndroid, a commit tells Java only its surface id. At the next mount dispatch, FabricUIManagerBinding::pullAndExecuteTransaction looks the surface's MountingCoordinator up in the binding's surface registry. A surface is not in the registry while it starts or stops:

• The transaction that stops a surface is never pulled, because stopSurfaceWithSurfaceHandler removes the surface from the registry before SurfaceHandler::stop() commits the empty tree. Mounting override delegates such as LayoutAnimations and C++ Animated never see those Deletes. With enableLayoutAnimationsOnAndroid, a layout animation that runs when its surface stops stays in flight, and FabricUIManager calls driveCxxAnimations on every frame until another surface pulls a transaction.
• A commit that a bridgeless surface makes before startSurfaceWithSurfaceHandler registers it is not mounted until the surface commits again.
• The legacy stopSurface(int) holds the registry lock across stop(). On the UI thread, the stop commit pulls inline and blocks on the same lock, so the app freezes.

iOS does not look the surface up. RCTMountingManager captures the MountingCoordinator that reported the commit in the block it dispatches to the main queue, and pulls from it.

This PR does the same. schedulerShouldRenderTransactions keeps the reporting MountingCoordinator until the UI thread pulls, and pullAndExecuteTransaction pulls from it. Java still receives only the surface id. The surface registry, the start and stop functions, reportMount and the push model are unchanged.

As on iOS, a surface stopped on the UI thread has its stop transaction pulled during the stop commit. When it stops on another thread, its shadow tree is destroyed and its coordinator revoked before the UI thread pulls, so that transaction is not mounted. A pull of a stopped surface can log "Executing commit after surface N was stopped!" when a mounting override delegate still returns a transaction; iOS pulls these the same way.

Only one coordinator per surface id is kept. Stopping and starting the same ReactSurface on the same host reuses its id. If the stopped surface's commit has not been pulled yet, the restarted surface's coordinator replaces it and a warning is logged. The replaced coordinator is revoked and would pull nothing. Java does not mount such a restarted surface anyway: MountingManager.startSurface logs "Called startSurface more than once" and keeps the stopped surface's manager.

Changelog:

[ANDROID] [FIXED] - Pull model: mount the commits of starting and stopping surfaces, and fix a freeze when the legacy stopSurface(int) runs on the UI thread

Test Plan:
New FabricMountingManagerInstrumentationTest cases drive a real FabricUIManagerBinding. The test helper runs no JavaScript and commits views to the surface's shadow tree itself. The pull-model cases run with the Props 2.0 flags and enablePreallocatedPropsDiffOnInsertAndroid.

Test	Without the fix	With the fix
native_stopSurfaceWithSurfaceHandler_pullModel_pullsDeletes	fails: no Deletes are pulled	passes
native_stopSurface_pullModel_pullsDeletes	hangs	passes
native_stopSurfaceWithSurfaceHandler_pushModel_pullsDeletes	passes	passes

The push case is a control. The hang stops the run, so the bridgeless stop test was also run without the legacy one. With the fix, the whole class passes.

A one-off check, not committed, covered the start: a bridgeless surface started on a background thread commits a view, and the start is held before startSurfaceWithSurfaceHandler registers the surface. On #59029 alone the view is not mounted while the start is held. With this PR it is mounted by the next frame.

The test class has no Gradle target in the open-source repository, so I ran it on an Android emulator through a local androidTest source set:

./gradlew :packages:react-native:ReactAndroid:assembleReleaseAndroidTest -PreactNativeArchitectures=arm64-v8a
adb install -r -t packages/react-native/ReactAndroid/build/outputs/apk/androidTest/release/ReactAndroid-release-androidTest.apk
adb shell am instrument -w -e class com.facebook.react.fabric.FabricMountingManagerInstrumentationTest \
  com.facebook.react.tests.gradle/androidx.test.runner.AndroidJUnitRunner

I also checked the effect in RNTester with a temporary hook, not part of this PR. The Playground starts a 60 s LayoutAnimation, then ReactDelegate.unloadApp() stops the surface on the UI thread while the activity stays resumed. Props 2.0 and enableLayoutAnimationsOnAndroid were on:

Run	After the stop
Pull, without this PR	no onAllAnimationsComplete; driveCxxAnimations keeps running on every frame
Pull, with this PR	onAllAnimationsComplete right after the stop, then no more driven frames
Push, with this PR	onAllAnimationsComplete right after the stop, as before

Pulled By:
bartlomiejbloniarz

bartlomiejbloniarz

Differential Revision: D124966432

[meta-codesync]

@bartlomiejbloniarz has imported this pull request. If you are a Meta employee, you can view this in D124966432.

[meta-codesync]

@bartlomiejbloniarz has exported this pull request. If you are a Meta employee, you can view the originating Diff in D124966432.