Preview the next exercise and count into each set during rests

The run flow's between-exercise rest now previews the next exercise — its
figure and a large "Coming up" name — and the hands-free narration gains a
per-second count-in ("<Exercise>, in 1, 2, 3, GO!") spoken before every set,
plus a "Coming up" announcement as a between-exercise rest begins. A new
Settings > Narration picker chooses the count-in cues, the setup/form read,
or both.

Also fixes spoken cues going silent after the first exercise in a flow split:
the stop-on-teardown moved from the per-exercise view (rebuilt on every
hand-off) to the run host, which stays mounted for the whole run. The audio
session now holds its duck across the per-second count so background music
doesn't pulse between words.

Claude-Session: https://claude.ai/code/session_01BQcEWmAPA78338QuEwRkAh

Workouts/Speech/SpeechAnnouncer.swift

@@ -26,16 +26,25 @@ final class SpeechAnnouncer: NSObject, AVSpeechSynthesizerDelegate {
 
     private let synthesizer = AVSpeechSynthesizer()
 
+    /// While true, `deactivateSession` won't hand audio focus back when an utterance ends —
+    /// so a per-second spoken countdown ("in 1… 2… 3… Go") stays ducked throughout instead
+    /// of pulsing other audio back to full in the gaps between counts. Cleared by the final
+    /// (non-holding) utterance of a run and by `stop`.
+    private var holdsSession = false
+
     override init() {
         super.init()
         synthesizer.delegate = self
     }
 
     /// Speak `text`, replacing anything currently playing. Empty/blank text is a no-op.
-    func speak(_ text: String) {
+    /// Pass `holdSession: true` for all but the last of a rapid back-to-back sequence so
+    /// other audio stays ducked between the utterances rather than un-ducking in each gap.
+    func speak(_ text: String, holdSession: Bool = false) {
         let trimmed = text.trimmingCharacters(in: .whitespacesAndNewlines)
         guard !trimmed.isEmpty else { return }
 
+        holdsSession = holdSession
         activateSession()
         // Replace, don't queue: a new exercise's cue supersedes the previous one.
         if synthesizer.isSpeaking {
@@ -56,6 +65,7 @@ final class SpeechAnnouncer: NSObject, AVSpeechSynthesizerDelegate {
 
     /// Stop immediately and hand audio focus back to other apps.
     func stop() {
+        holdsSession = false
         if synthesizer.isSpeaking {
             synthesizer.stopSpeaking(at: .immediate)
         }
@@ -73,10 +83,11 @@ final class SpeechAnnouncer: NSObject, AVSpeechSynthesizerDelegate {
         try? session.setActive(true)
     }
 
-    /// Return other audio to full volume. Skipped while something is still speaking, so a
-    /// cancel-then-speak (new exercise) doesn't briefly un-duck between utterances.
+    /// Return other audio to full volume. Skipped while something is still speaking, or while
+    /// a countdown holds the session, so a cancel-then-speak (new exercise) or a per-second
+    /// count doesn't briefly un-duck between utterances.
     private func deactivateSession() {
-        guard !synthesizer.isSpeaking else { return }
+        guard !synthesizer.isSpeaking, !holdsSession else { return }
         try? AVAudioSession.sharedInstance().setActive(false, options: [.notifyOthersOnDeactivation])
     }