This PR introduces a bunch of UI/UX improvments as well as some refactoring to the existing behaviours.

Among the UI improvments, it adds:
- a new `AudioVisualizer` component which can be sized dynamically,
- uses the standard `blockquote` block for rendering transcriptions,
- Completely change the controls of the component on mobile to make it more reactive and usable. Specifically by making the navigation and configuration controls use the `Dropdown` component when on mobile,
- The transcription visuals are now divided in three main steps: listening step to add the "AI is listening..." text, delta step where the transcription bits are added to the current paragraph, and the completed step which happens on longer breaks to create a new paragraph.

As for the refactoring:
- Removal of all the embedded states, keeping them internal to the component and using standard browser API's to observe changes (i.e. the `MutationObserver` to notify changes on the summary),
- The component now relies on the descendents to apply changes during the transcription process, getting rid of the ids,
- Splitting of the `VADAudioRecorder` into separate classes: `AudioProcessor` and `RealtimeClient`. Performing audio recording, and WebSocket operations respectively,
- Making the audio processing, and websocket operation rely on singletons to avoid issues when the editor has to be reloaded in case of synchronisation issues,
- The audio processing was also simplified, and merged with the `PCM16AudioProcessor` using two new events: "vad" for client voice detection, and "audio" for sending the regular audio information. The buffer size was tweaked from 6000 to 2400, to prevent early cut-off of the audio by the server_vad. All the pre-filtering of the audio was also removed which improves the overall quality of the transcript.

task-5153836