Fix pyarrow version which causes winml eval crash

[zhenchaoni]

Fixes #749

PR Description

Fix: winml eval crash when --device is not specified

Problem

Running winml eval without an explicit --device flag crashes the process with an access violation (0xC0000005) during dataset loading.

Root Cause

The crash is triggered by a DLL incompatibility between pyarrow 24.x and the WinML/ONNX Runtime EP DLLs:

1. When --device is omitted, winml eval resolves the device and preloads the EP DLLs (via _preload_bundled_onnxruntime_dll) before importing the datasets package.
2. Importing datasets then loads pyarrow's C extension on top of the already-loaded EP DLLs.
3. pyarrow 24.x ships a native binary that conflicts with the WinML EP DLLs in this load order, causing the access violation.

This was only reproducible with pyarrow 24.x; versions ≤ 23 load cleanly regardless of order.

Fix

Cap pyarrow to >=21,<24 in pyproject.toml. This stays within the range required by datasets (which needs pyarrow>=21), so no other dependencies are affected.

The existing comment above the pin already documents the crash details and the conditions for revisiting the cap (re-test 24.x patches / 25.x before relaxing).

Validation

• uv run winml eval -m microsoft/resnet-50 --samples 100 now completes without --device specified.
• No change to runtime behavior for users who already pinned pyarrow < 24.