normalize media pipeline at client boundary

- AudioParams.framing field: client declares "raw" or "adts"
- Client strips ADTS from audio before sending (strip_adts)
- Client does H.264 NAL inspection for keyframe detection (h264_is_keyframe)
- Server uses declared sample_rate/channels for ADTS synthesis instead of hardcoded 48kHz/stereo
- Server gates ADTS wrapping on framing field instead of per-packet sniffing

New backends only need to pipe output to demux_and_send() — server and Python unchanged.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

media/server/src/main.rs

@@ -98,9 +98,9 @@ async fn handle_client(
                 info!("control: {ctrl:?}");
 
                 match ctrl {
-                    ControlMessage::SessionStart { id, video, .. } => {
+                    ControlMessage::SessionStart { id, video, audio } => {
                         let s = tokio::task::block_in_place(|| {
-                            Session::start(&id, &sessions_dir, video.fps)
+                            Session::start(&id, &sessions_dir, video.fps, &audio)
                         })?;
                         session = Some(s);
                     }