Update eval_squad to use API of latest optimum (#17918)

[TedThemistokleous]

Update eval_squad with latest optimum.

Tested with:

• optimum 1.13.1
• transformers 4.31.0
• onnxruntime-gpu 1.16.0
• onnx 1.14.1
• datasets 2.14.5
• evaluate 0.4.0
• torch version 2.2.0.dev20230920+cu121

Example output in A100:

{'exact': 86.66035950804162, 'f1': 92.99622739711005, 'total': 10570, 'HasAns_exact': 86.66035950804162, 'HasAns_f1': 92.99622739711005, 'HasAns_total': 10570, 'best_exact': 86.66035950804162, 'best_exact_thresh': 0.9998456239700317, 'best_f1': 92.9962273971104, 'best_f1_thresh': 0.9998456239700317, 'total_time_in_seconds': 84.74025378189981, 'samples_per_second': 124.73410838731417, 'latency_in_seconds': 0.008017053337928081, 'provider': 'CUDAExecutionProvider', 'disable_fused_attention': False, 'pretrained_model_name':
'bert-large-uncased-whole-word-masking-finetuned-squad', 'onnx_path': './bert-large-uncased-whole-word-masking-finetuned-squad/optimized_model.onnx', 'batch_size': 1, 'sequence_length': 384, 'use_io_binding': True}

Description

Cherry pick from Microsoft/Onnxruntime for upstream changes to eval_squad.py

Motivation and Context

Without this backport we're broken for migx_onnxrt_bert_distilled_benchmarks due to missing pieces in the script. Testing was originally done upstream and it appears this change hasn't been backported.

[TedThemistokleous]

Output using MIGraphX EP on Navi31

Evaluation done in 15.5 seconds
{'exact': 100.0, 'f1': 100.0, 'total': 2, 'HasAns_exact': 100.0, 'HasAns_f1': 100.0, 'HasAns_total': 2, 'best_exact': 100.0, 'best_exact_thresh': 0.7958137392997742, 'best_f1': 100.0, 'best_f1_thresh': 0.7958137392997742, 'total_time_in_seconds': 14.235455489950255, 'samples_per_second': 0.1404942751155333, 'latency_in_seconds': 7.117727744975127, 'provider': 'MIGraphXExecutionProvider', 'disable_fused_attention': False, 'pretrained_model_name': 'distilbert-base-cased-distilled-squad', 'onnx_path': './distilbert-base-cased-distilled-squad/model.onnx', 'batch_size': 1, 'sequence_length': 384, 'use_io_binding': False}

[groenenboomj]

This PR has been working upstream and we usually catch this in our remote syncs but it does look like this may have been missed. For the sake of normalization I think it's good to merge this and spot check for other related missed bits.

[TedThemistokleous]

This PR has been working upstream and we usually catch this in our remote syncs but it does look like this may have been missed. For the sake of normalization I think it's good to merge this and spot check for other related missed bits.

I have checked it and it seems to work properly now for eval_squad.py.

[TedThemistokleous]

@groenenboomj is there anywhere else I should be backporting this too? rocm6.0_internal as well?