Fix quantization P0 bugs

[zhenchaoni]

PR Title

fix(quantize/eval): validate --precision, stream default image dataset, support pixel_mask padding

PR Description

Summary

Three independent fixes bundled together — two are user-reported bugs, one is a side-fix found while validating object detection.

Changes

1. Reject invalid --precision values (fixes #555)

Previously, winml quantize --precision banana (or any unrecognized string) silently fell through to uint8/uint8 defaults and reported "Success!", masking user typos like w4a16 or fp64. _resolve_quant_types now raises click.BadParameter for any precision that is not None, "auto" (case-insensitive), or a value accepted by is_quantized_precision. The error message lists the accepted forms.

2. Stream the default image calibration dataset (fixes #433)

The default calibration dataset (timm/mini-imagenet) was downloading the entire training split (~5 GB) just to pull 10 calibration samples. ImageDataset._get_default_dataset now sets streaming=True by default, and the shared _load_and_sample helper takes the streaming path when streaming is enabled and max_samples is set (streaming without a sample cap would be strictly worse than a bulk download, so it falls back automatically). ObjectDetectionDataset reuses the same helper, eliminating duplicated load/sample logic.

Verified: with the HF cache cleared, running winml quantize -m resnet-50 --task image-classification --model-name microsoft/resnet-50 --samples 10 leaves datasets--timm--mini-imagenet at 0 MB on disk.

3. Support pixel_mask input for DETR-family object detection

Some object detection exports (DETR, etc.) declare a pixel_mask input alongside pixel_values, which the HF image processor only populates meaningfully when padding to a fixed canvas. WinMLObjectDetectionEvaluator.prepare_pipeline now detects this case and switches the processor to aspect-preserving resize (shortest_edge/longest_edge) + pad_size + do_pad=True. Models without pixel_mask keep the previous exact-resize, no-padding behavior.

Tests

• TestQuantizePrecisionValidation — parametrized over banana, w4a16, int4, fp64; asserts quantize_onnx is never invoked.
• TestPreparePipeline::test_pixel_mask_path_enables_padding and test_no_pixel_mask_path_disables_padding — cover both branches with a sentinel to prove the wrong branch did not run. Uses monkeypatch for the parent-class prepare_pipeline stub so the override is reverted on teardown.

All 84 affected unit tests pass.

Manual verification

Issue	Verification
#555	winml quantize -p banana → BadParameter: 'banana' is not a supported quantization precision. Expected: auto | int8 | int16 | w{x}a{y} ...
#433	Cleared HF cache → ran quantize → datasets--timm--mini-imagenet = 0.0 MB on disk; only 10 examples flowed through Map/Aligning the labels
pixel_mask	Covered by new unit tests

[timenick]

Five findings — none blocking, but worth addressing before merge. The biggest one is that the new error message redirects users to --help, but --help doesn''t yet mention auto as an accepted value, so users hitting the BadParameter still won''t know how to fix it.