Fix InputMethodServiceInternal lifetime

This is a follow up CL to my recent CL [1], which aimed to consolidate
dependencies on InputMethodService from RemoteInputConnection.

One thing I overlooked was that InputMethodService did not retain a
strong reference to InputMethodServiceInternal object that I
introduced there, which means that the newly instantiated
InputMethodServiceInternal could be garbage-collected even before
InputMethodService became unreachable.

As a result, the following InputConnection APIs could start failing at
some point and would never work without restarting the IME, which is
likely to be the root cause of the recent flakiness in
InputConnectionEndToEndTest discussed in Bug 193822485.

  * InputConnection#requestCursorUpdates()
  * InputConnection#commitContent()

Also, the following functionality might have stopped working at some
point.

  * InputMethodService#dump()
  * InputConnection tracing in IME tracing
  * Updating IME switching order based on user inputs

With this CL, the InputMethodServiceInternal instance is guaranteed to
be instantiated once and remain to be alive until
AbstractInputMethodService becomes unreachable.

 [1]: I2aeeeacd27195ce10059d6590e098a4a969e774d
      a975bfc4

Fix: 193822485
Fix: 199288389
Test: atest CtsInputMethodTestCases
Test: Manually verified with Imege Keyboard sample IME/app
Change-Id: I7ee0270bad94ed606048bfaf76e8545336715d83