Fix input validation issues and update multiple-choice options handling

[apsonawane]

This pull request introduces improvements to the handling of multiple-choice questions and options extraction in the data preprocessing and evaluation pipeline, as well as a minor enhancement to ONNX model input handling. The main changes include switching from numeric to letter-based multiple-choice options, extracting the selected letter during evaluation, and supporting an additional input name in ONNX graph surgeries.

Multiple-choice options handling:

• Updated pre_process_data.py to format multiple-choice options using letters (A, B, C, ...) instead of numbers, and added a new field extract_option_letter to indicate when letter extraction is needed. [1] [2]
• Modified olive_evaluator.py to extract the selected letter from model responses for multiple-choice questions, in addition to the previous digit extraction. This ensures compatibility with the new letter-based formatting. [1] [2] [3]

ONNX model input handling:

• Enhanced the TieWordEmbeddings class in graph_surgeries.py to support both input_embeds and inputs_embeds as valid input names, improving compatibility with different model variants.

[Copilot]

Pull request overview

This pull request updates the ONNX TieWordEmbeddings graph surgery to be more compatible with exported model interfaces by recognizing "inputs_embeds" as an additional valid input name (alongside "input_ids" and "input_embeds").

Changes:

• Extend TieWordEmbeddings.__call__ input name detection to include "inputs_embeds".
• Update the inline comment to reflect the supported input-name variants.