Reverts d5aff71 (#11948) and implements an alternative solution.
Even with the previous fix applied, we were calling `renderFinished()` directly when the render was canceled. This led to the following scenario (all on the main thread):
- `Map::renderStill()` callback is called, `uv_async_send` schedules a call to `renderFinished()`.
- `NodeMap::cancel()` destroys the map object, and calls `renderFinished()` directly, which cleans up the request
- `NodeMap::startRender()` is called to schedule a _new_ render, which means that a new callback will be set
- The `uv_async_t` fires and calls `renderFinished` again:
- We don't return early as intended by the fix in #11948 since a new callback/`req` object is now available.
- Since neither error nor image has been set, we'll see a `Didn't get an image` error message.
- We call the callback with the error. In scenario where we're reusing Map objects, we will start a new render, which will now trigger the `Map is currently rendering an image` error in `Map::renderStill`, and the map object enters a circle in which it will now fail repeatedly with this error.
This commit fixes the issue by completely removing the existing `uv_async_t` handle that could have a scheduled call, and replacing it with a fresh instance of an `uv_async_t` handle.
f3b5902f05