Use the rc2 diffusers wheel for enhancement constants

Same LTX-2.4 system prompts, but GEMMA4_PROMPT_ENHANCEMENT_CONFIG raises
no_repeat_ngram_size from 3 to 5. At 3 the enhancer occasionally emitted odd tokens
mid-caption (a Devanagari word, a dropped space) as it was forced off repeated n-grams.

app.py

@@ -14,7 +14,10 @@ from huggingface_hub import hf_hub_download
 HF_TOKEN = os.environ.get("HF_TOKEN")
 _whl = hf_hub_download(
     "diffusers-internal-dev/ltx24-wheels",
-    "canon/diffusers-0.40.0.dev0-py3-none-any.whl",
+    # rc2 wheel (PR head): identical LTX-2.4 system prompts, but GEMMA4_PROMPT_ENHANCEMENT_CONFIG
+    # raises no_repeat_ngram_size 3 -> 5. At 3, long descriptive captions get pushed off common
+    # phrasing and occasionally emit odd tokens (observed a Devanagari word spliced mid-caption).
+    "rc2/diffusers-0.40.0.dev0-py3-none-any.whl",
     repo_type="dataset",
     token=HF_TOKEN,
 )