Fix datasets task 7309: pin pyarrow < 20.0 for geoparquet compatibility (#43)

PyArrow 20.0 changed default string type mapping, causing geoparquet
geometry columns to report as 'large_string' instead of 'string'. The
pre-existing test_parquet_read_geoparquet asserts dtype == 'string',
breaking all feature pairs for this task.

Pin pyarrow to >=17.0.0,<20.0.0 to keep a modern version while avoiding
the large_string default change. The Docker image needs to be rebuilt.

Author: AlienKevin

dataset/huggingface_datasets_task/task7309/Dockerfile

@@ -20,7 +20,7 @@ RUN git clone https://github.com/huggingface/datasets.git repo && \
 # Set up Python environment and pre-install dependencies
 WORKDIR /workspace/repo
 RUN uv pip install --system -e .
-RUN uv pip uninstall --system pyarrow && uv pip install --system "pyarrow==20.0.0"
+RUN uv pip uninstall --system pyarrow && uv pip install --system "pyarrow>=17.0.0,<20.0.0"
 # Install tensorflow, explicitly excluding tensorflow-macos (macOS-only package)
 RUN uv pip install --system "tensorflow>=2.16.0,<2.17.0" || uv pip install --system "tensorflow==2.16.2"
 RUN uv pip install --system torch jax