Add MTEBEvaluator for embedding model evaluation

[natke]

Summary

Add a built-in MTEB (Massive Text Embedding Benchmark) evaluator to Olive, following the same architecture as LMEvaluator/lmeval_ort.py.

Changes

olive/evaluator/olive_evaluator.py

• New MTEBEvaluator class registered in the evaluator registry
• Supports three model classes: hf (SentenceTransformer), ort (ONNX via ORT), ortgenai (GenAI with hidden_states)
• Auto-detects model class from handler type, including GenAI detection via genai_config.json

olive/evaluator/mteb_ort.py (new)

• MTEBOnnxBase: abstract base implementing MTEB's EncoderProtocol with mean pooling
• MTEBORTEvaluator: wraps plain ONNX models via ort.InferenceSession
• MTEBORTGenAIEvaluator: wraps GenAI models using og.Generator + hidden_states output

Usage

"evaluators": {
    "mteb": {
        "type": "MTEBEvaluator",
        "tasks": ["STS17"],
        "batch_size": 32
    }
},
"evaluator": "mteb"

Testing

Tested end-to-end with Qwen3-Embedding-0.6B CPU recipe — ModelBuilder export + MTEB STS17 evaluation on both input HF model and exported GenAI model.

[Copilot]

Pull request overview

Adds a new built-in evaluator to Olive for running MTEB (Massive Text Embedding Benchmark) embedding evaluations, following the existing evaluator registry pattern used for LM evaluation.

Changes:

• Registers a new MTEBEvaluator in olive_evaluator.py with support for HuggingFace, plain ONNX Runtime, and ORT GenAI style models.
• Introduces olive/evaluator/mteb_ort.py with ONNX Runtime and ORT GenAI adapter classes implementing an MTEB-compatible encoder interface.
• Converts MTEB task results into Olive MetricResult structures for reporting.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 4 comments.

File	Description
olive/evaluator/olive_evaluator.py	Adds MTEBEvaluator and wiring to select the correct backend and translate MTEB outputs into Olive metrics.
olive/evaluator/mteb_ort.py	Adds ONNX Runtime and ORT GenAI model wrappers to satisfy MTEB encoder requirements (tokenization, pooling, similarity).