fix(eval): respect image-to-text pipeline call contracts

[Yue Sun (KayMKM)]

Summary

• Supply an empty text prompt only when the active pipeline explicitly declares that parameter. Legacy ImageToTextPipeline receives only the image; multimodal ImageTextToTextPipeline still receives its required empty prompt.
• Raise an actionable error with the original cause when pipeline failures leave no evaluated predictions. Preserve partial results and empty-dataset behavior.
• Cover both real Hugging Face pipeline call contracts using generated images/references, and cover all-sample failures. Document the behavior.
• Keep the existing E2E CER/CIDEr thresholds, sample-count assertion, and no-skips assertion unchanged.

Background

The daily E2E run 156998365 fails TestEvalPerTask::test_image_to_text_fp16 on QNN and AMD. Every sample raises ImageToTextPipeline._sanitize_parameters() got an unexpected keyword argument 'text', and the evaluator returns cer=None after skipping them all. Fix the evaluator rather than relaxing the E2E assertions.
The independent AMD EP-discovery failures and OV agent disconnect are out of scope.

Validation

• 46 focused pytest cases passed, covering evaluator behavior, text metrics, task mapping and evaluator/CLI error propagation.
• Ruff passed for the changed Python files.
• Full model/hardware E2E is left for manual execution on a host where the hardware-gated case executes rather than skips:

uv run --no-sync pytest tests\e2e\test_eval_e2e.py::TestEvalPerTask::test_image_to_text_fp16 -m e2e --timeout=1800