Point the missing-checkpoint error at the model repo

scripts/fetch_models.py does not exist in either published repo, so anyone
who hit this error was sent to a dead command. The weights arrive through
from_pretrained now.

whisper_decoder.py

@@ -136,8 +136,10 @@ class Whisper:
         if not (model_dir / "config.json").exists():
             raise FileNotFoundError(
                 f"{model_dir} is not a Whisper checkpoint directory "
-                "(no config.json). Fetch the weights with "
-                "python scripts/fetch_models.py"
+                "(no config.json). The weights come from the model repo, so "
+                "load through it rather than pointing here by hand:\n"
+                "    AutoModel.from_pretrained("
+                "'Daporte/prosody2text-weights', trust_remote_code=True)"
             )
 
         if device is None: