Fix zero-shot calibration defaults

[DingmaomaoBJTU]

Summary

Fix clean-cache QNN zero-shot-classification E2E instability by making zero-shot calibration task-specific instead of reusing the generic text-classification calibration default.

Root cause

Static quantization creates a default calibration reader when callers do not provide custom calibration data. Before this change, zero-shot-classification was mapped to TextDataset, whose default dataset is glue/mrpc.

On clean agents with the current HF stack, load_dataset("glue", "mrpc") can fail while resolving the repo metadata path:

Invalid HF URI 'hf://datasets/glue@.../.huggingface.yaml'. Repository id must be 'namespace/name', got 'glue'.

That failure caused calibration to fall back to RandomDataset. The random token inputs are not representative for an NLI-style zero-shot classifier, so the QNN quantized artifact could produce a low F1 score in the E2E test.

Fix

• Add ZeroShotClassificationDataset for calibration.
• Map zero-shot-classification to this specialized dataset instead of TextDataset.
• Use the same built-in zero-shot defaults as evaluation: fancyzhx/ag_news, test split, candidate labels World,Sports,Business,Sci/Tech, and hypothesis template This text is about {}.
• Expand each source text into premise/hypothesis pairs before tokenization, matching the input distribution expected by zero-shot/NLI models.
• Leave other task mappings and generic TextDataset behavior unchanged.

Validation

• uv run ruff check --fix src\winml\modelkit\datasets\zero_shot_classification.py src\winml\modelkit\datasets\__init__.py tests\integration\datasets\test_zero_shot_classification.py
• uv run pytest tests\integration\datasets\test_zero_shot_classification.py tests\unit\datasets\test_random_dataset.py -vv
• Cleared glue HF cache and the cross-encoder/nli-deberta-v3-small WinML artifact cache, then ran:
  uv run pytest -m e2e tests\e2e\test_eval_e2e.py::TestEvalPerTask::test_zero_shot_classification -vv -s --tb=short