Add semaphore to limit concurrent prefetch operations to 1 at a time.
This helps prevent potential race conditions in native audio libraries
(CPAL/PipeWire/ALSA) that can cause malloc corruption.
Also:
- Reduce QOBUZ_PREFETCH_COUNT from 3 to 2
- Reduce PREFETCH_LOOKAHEAD from 15 to 10
- Add 50ms delay before cache insertion
Addresses #38 - random crashes during prefetch