* Replace hard break of Layer.find/All with findAllAnnotations
* Deprecate findAll

* Revert "Revert "Rewrite MouseTracker's tracking and notifying algorithm (#42031)" (#42478)"

This reverts commit eede7929.

* Fix tests

* Redesigns the algorithm with which MouseTracker keeps track of states and dispatches events. 

* Make sendMouseNotifications private

This fixes the mouse hover code to not schedule frames with every mouse move.

Before this, it would schedule a post frame callback, and then schedule a frame immediately, even if there was nothing that needed to be updated. Now it will schedule checks for mouse position updates synchronously, unless there's a new annotation, and skip scheduling a new frame in all cases. It has to be async in the case of a new annotation (i.e. a new MouseRegion is added), since when the annotation is added, it hasn't yet painted, and it can't hit test against the new layer until after the paint, so in that case it schedules a post frame callback, but since it's already building a frame when it does that, it doesn't need to schedule a frame.

The code also used to do mouse position checks for all mice if only one mouse changed position. I fixed this part too, so that it will only check position for the mouse that changed.

This PR solves two problems: currently, the onExit is called for a mouse pointer the moment the removal message is received, except that by the time it actually calls it, there is no _lastEvent for it in the mouse tracker (it's already been removed), resulting in an event being passed to the onExit that contains nulls for the position. Also, removePointer events don't actually get created with a position, although they easily could be, so that even the the _lastEvent in the mouse tracker were still populated, it would still give a null position and delta.

This PR adds support for the position and delta in a PointerRemovedEvent, and populates them. In addition, when a remove event is received, it doesn't actually remove the pointer until the mouse position check that gets scheduled actually happens.

This changes Listener to trigger enter/move/exit in all Listeners below the pointer, not just the leaf region (the first region hit). This is because we need to allow listeners to be nested so that, say, a widget that handles changing color on hover, but also is wrapped in a Tooltip (that handles hover) can trigger both actions, not just one.

To that end, I added a findAll to Layer, similar to the existing find method that was previously used. It returns an iterator over annotated layers which match the given data type.

Since the findAll is implemented as returning an Iterable (and is sync*), I re-implemented the find routines as just returning the first result from findAll, since that should be just as efficient, and would then prevent duplication in the implementation.

I recently added some code to keep hover events from being propagated when a mouse wasn't attached. While that works, there are times when it can fire callbacks during the building of other components, since they can now be called from detach/attach. This is not ideal, since it will assert then. This changes the code so that it won't update the annotations during attach/detach, but also won't push the annotation layer unless a mouse is connected, achieving the same result as before, but with better semantics.

The basic problem is that in the detach for RenderPointerListener, it would detach the annotation, which could cause onExit to be called on the annotation, since the widget was disappearing under the mouse, and thus needs to receive an onExit, but that onExit might be (and probably will be) calling setState, which marks the owning widget as needing to be built, sometimes when it already has been.

The fix creates a new _ListenerElement that overrides activate and deactivate in order to tell the render object ahead of the detach that it might be detached, and so the onExit gets called before the detach instead of during it.

In addition, I now avoid scheduling more than one check for mouse positions per frame.

This is a re-land of #31561, after fixing performance regressions.

Added change listening to the MouseTracker so that the Listener and tooltip can react to whether or not a mouse is connected at all. Added a change check to make sure Listener only repaints when something changed.

Fixes #22817

* Allow mouse hover to only respond to some mouse events but not all.

Adds a paint after detaching/attaching hover annotations to keep the annotation layers in sync with the annotations attached to the mouse detector.

Fixes #30744

Avoids an implicit downcast + crash in MouseTracker, since the last
pointer event can be a Move or Down event, in addition to Hover.

* Fix MouseTracker annotation leak

Map's remove method is typed `remove(Object key)`, which can cause bugs.

Modified the existing test to check if the annotation has been removed
from MouseTracker.

This implements support for hovering mouse pointers, so that mice connected to Android devices, and ChromeOS devices running Android apps will work properly.

It teaches flutter_test about hover events, which required changing how they are created and used.

Also modifies AnnotatedRegion to allow a region that can be located someplace other than just the origin.

Along with tests for all of the above.

Fixes #5504