Cache voice conditionals and add FP16 autocast

Voice conditionals (s3tokenizer + voice encoder + mel embeddings) are
expensive to compute but depend only on the reference audio, not the
text. Previously they ran on every synthesis chunk — 3x wasted work for
a 3-chunk request. Now computed once at startup and reused.

Also wrap generate() in torch.amp.autocast(float16) for ~2x speedup on
all model computation (T3 LLM, S3Gen CFM, HiFiGAN vocoder).

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

main.py

@@ -19,8 +19,18 @@ logger = logging.getLogger(__name__)
 
 
 def _warmup(voices: dict) -> None:
-    """Run one synthesis to populate MIOpen's in-memory kernel cache."""
+    """Pre-compute voice conditionals and populate MIOpen's kernel cache."""
     from wyoming_voices import resolve_voice
+
+    # Pre-compute conditionals for all discovered voices so the first real
+    # request doesn't pay the s3tokenizer + voice encoder cost.
+    for name, path in voices.items():
+        try:
+            engine.prepare_voice(path)
+        except Exception:
+            logger.warning(f"Failed to prepare voice '{name}' (non-fatal)", exc_info=True)
+
+    # Synthesis warmup to populate MIOpen's in-memory kernel cache.
     audio_prompt = resolve_voice(None, voices) if voices else None
     logger.info("Running warmup synthesis...")
     try: