Fix streaming audio startup de-click

[ivan-digital]

Summary

Add configurable startup de-click handling to StreamingAudioPlayer without adding default playback latency for existing model integrations.

The default remains conservative: no startup pre-roll and a short 5 ms first-chunk fade. Clients that need stronger protection for neural codec startup transients can opt into a short zero pre-roll and longer fade-in through startupPrerollDuration and startupFadeInDuration.

The PR also exposes underflowCount so streaming clients can distinguish actual mid-phrase playback starvation from normal silence after generation completes.

Why

Some neural codec TTS streams can begin with a non-zero or low-frequency transient. Starting the audio engine directly on that first sample can produce an audible dull click/pop at phrase start. Making the stronger de-click behavior configurable lets affected clients fix that startup artifact without changing timing for other models that already start cleanly.

Regression Safety

• Default startupPrerollDuration is 0, so existing playback paths do not gain extra startup latency.
• Default startupFadeInDuration is 0.005, matching the previous short first-chunk fade window.
• De-clicking is only applied to the first audible chunk of a generation.
• Underflow accounting ignores normal silence after generation completion.

Validation

• swift test --filter StreamingAudioPlayerTests
• 25 tests passed