Add continuous batching support for Qwen3 TTS to the server

[lucasnewman]

This implements vLLM Omni-style stepwise continuous batching for the Qwen3 TTS model (which is currently the only model we already have batch support for). I get nice parallelism when using 8 concurrent requests across 30 requests sent 300ms apart on my M5 Max:

Throughput
  Requests: 30/30 succeeded
  Client wall time: 21.70s
  Service window: 21.70s
  Requests/sec: 1.38
  Input chars/sec: 115.0
  Response bytes/sec: 378.02 KiB/s
  Audio throughput: 8.06x realtime (174.96s audio)
  End-to-end RTF: 0.12
  Parallelism factor: 7.07x

Latency
  min/mean/median/p95/max: 2.94s / 5.11s / 5.30s / 5.99s / 6.11s

Audio
  Decoded files: 30
  Sample rates: 24000
  Total samples: 4199040

It currently doesn't support streaming since the patch is already somewhat large and I want to stage in the complexity, but I'll look at both that and adding batch support for other models as follow-ons.