Pin numarkdown-ocr.py to stable vLLM (>=0.15.1) + datasets>=4.0.0

Qwen2.5-VL architecture is in stable vLLM, no nightly needed.
Pinning datasets>=4.0.0 fixes pyarrow compatibility.

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

numarkdown-ocr.py

@@ -1,10 +1,10 @@
 # /// script
 # requires-python = ">=3.11"
 # dependencies = [
-#     "datasets",
+#     "datasets>=4.0.0",
 #     "huggingface-hub",
 #     "pillow",
-#     "vllm",
+#     "vllm>=0.15.1",
 #     "tqdm",
 #     "toolz",
 #     "torch",  # Added for CUDA check