Add startup warmup and make fp16 autocast fault-tolerant

Warmup: run a synthesis before accepting Wyoming connections so MIOpen
benchmarks and caches all conv layer shapes. Without this, the first HA
request triggers hundreds of benchmark runs and times out.

fp16: wrap in try/except so a failed autocast retries in fp32 rather
than dropping the request silently.

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

engine.py

@@ -122,8 +122,13 @@ def synthesize(
         kwargs["exaggeration"] = exaggeration
         kwargs["cfg_weight"] = cfg_weight
 
-    with torch.inference_mode(), torch.autocast("cuda", dtype=torch.float16):
-        wav = chatterbox_model.generate(text=text, **kwargs)
+    try:
+        with torch.inference_mode(), torch.autocast("cuda", dtype=torch.float16, enabled=torch.cuda.is_available()):
+            wav = chatterbox_model.generate(text=text, **kwargs)
+    except Exception:
+        logger.warning("fp16 autocast failed, retrying in fp32", exc_info=True)
+        with torch.inference_mode():
+            wav = chatterbox_model.generate(text=text, **kwargs)
 
     if torch.cuda.is_available():
         torch.cuda.synchronize()