Shape ids are local to one presentation, and decoded audio was cached by the
address of a buffer the cache did not keep. Switching decks, closing the last
one, or replacing a clip could keep the previous sound. Playback now stops
when the document changes or closes, a new source replaces the voice's audio,
and the decode and duration caches hold the bytes they are keyed by.
Co-authored-by: Cursor <cursoragent@cursor.com>
