[code-new-bie]

Each audio renderer creates its codec and AudioTrack before it reads its
queue, and that took about 1.2 s on a BYD head unit. The queue held 64
packets and dropped the newest when full. A wired LPCM packet holds about
7 ms of audio, so 64 packets covered under half a second and the rest of the
start-up audio was dropped: the beginning of every wired stream was cut.

The queue is now bounded by the audio it holds, 2 s, as jitter buffers such
as GStreamer's rtpjitterbuffer and WebRTC's NetEq are. When it is over budget
the oldest audio goes, so a stalled output catches up on the newest audio and
latency stays bounded. Packet durations come from the codec: LPCM from the
payload size, AAC-LC as 1024 frames, Opus from the TOC byte (RFC 6716,
section 3.1). Items without a duration are never dropped, and a cap of 2048
items guards against packets that report none.

AudioPacketQueueTest covers wired start-up audio fitting without drops, AAC,
dropping the oldest audio, a blocking take and the per-codec durations.