Support HY-MT evaluation workflows

[hanbitmyths]

Summary

• Update lm-eval integration to preserve recipe bootstrap settings.
• Improve ORT GenAI evaluation provider handling and runtime GenAI search options.
• Keep model-load-only Hugging Face kwargs out of generation config loading.
• Add accelerator normalization coverage for WebGPU export-only and runtime-required paths.

Validation

• Ran py_compile for olive/evaluator/lmeval_ort.py, olive/evaluator/olive_evaluator.py, olive/model/handler/mixin/hf.py, and test/hardware/test_accelerator.py.
• Ran git diff --check for the touched Olive files.
• VS Code diagnostics reported no errors for touched accelerator/evaluator files.
• Targeted pytest was attempted but blocked by missing pydantic in this environment.

[Copilot]

Pull request overview

This PR enables HY-MT evaluation workflows by tightening lm-eval/ORT-GenAI integration: it preserves recipe-provided bootstrap_iters, normalizes ORT GenAI provider names and mirrors the exported past_present_share_buffer setting at runtime, prevents model-loading-only kwargs from leaking into generation config loading, and extends accelerator tests with WebGPU coverage.

Changes:

• LMEvaluator now forwards a configurable bootstrap_iters to lm_eval.simple_evaluate.
• LMEvalORTGenAIEvaluator maps Olive EP names to ORT GenAI provider names, reads past_present_share_buffer from genai_config.json, and renames device to _device.
• HfMixin.get_hf_generation_config filters out torch_dtype/device_map/max_memory/quantization_config before calling get_generation_config.
• New accelerator normalization tests for WebGPU in both skip-EP-check and runtime-required paths.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

File	Description
olive/evaluator/olive_evaluator.py	Adds bootstrap_iters plumbing into LMEvaluator and simple_evaluate.
olive/evaluator/lmeval_ort.py	Adds ORT GenAI provider name normalization, runtime past_present_share_buffer, and renames self.device to self._device.
olive/model/handler/mixin/hf.py	Excludes model-load-only kwargs when loading HF generation config.
test/hardware/test_accelerator.py	Adds WebGPU coverage for accelerator normalization skip/required paths.