Speech: Concurrent recognition service

Currently, each recognition service implementation can run a single
recognition session at one time. If another client tried to start a
session, it would fail with SpeechRecognizer.ERROR_RECOGNIZER_BUSY.
With these changes, each recognition service implementation can run
multiple recognition sessions at one time.

Service implementations can control the level of concurrency by
overriding the newly added `getMaxConcurrentSessionsCount` method.
The default value is 1, meaning no concurrency.

A CTS testing these changes is in the same topic as this CL.

Bug: 261895777
Test: Manual, using the intent-starter test app
Change-Id: Ia64f4bb6aeff89942b3cfd2efb3520b5b3ac5ae7