Implement head_qa benchmark for inference evaluation pipeline

[boysack]

Overview

This PR officially integrates the head_qa (HEAlthcare Dataset) benchmark into the evaluation pipeline. This enables the evaluation of Meditron models against multi-choice questions derived from the Spanish healthcare system specialization exams.

Core Implementation

• Benchmark Integration: Created the HeadQA class in benchmarks.py, enabling the downloading, parsing, and execution of the dataset within the standard inference.py and evaluate.py workflow.
• Custom Preprocessing: Implemented the custom_preprocessing static method to accurately map head_qa's nested question/answer structure into the standard A/B/C/D/E multiple-choice format expected by the Meditron evaluator.
• Dataset Routing Workaround: Bypassed current upstream Hugging Face deprecation blockers on the official repo by pointing the pipeline to a stable mirror (boysack/head_qa). This ensures the evaluation runs seamlessly out-of-the-box.
• Native Configuration: Added meditron-7b and meditron-7b-awq Hugging Face paths directly into the master checkpoints dictionary in inference_pipeline.sh so the benchmark can be run natively without passing manual URIs.

Other fixes

• vLLM & Memory Optimizations: Enabled AWQ quantized models (meditron-7B-AWQ), bypassed FlashInfer warmup spikes (enforce_eager=True), capped VRAM utilization at 0.8, and removed the hardcoded tensor_parallel_size.
• vLLM & SamplingParams Deprecation: Removed deprecated parameters (best_of, use_beam_search).
• Framework Maintenance: Updated Hugging Face caching path to namespace___resource (must be updated in the other benchmarks), patched the missing instructions.json template, and fixed evaluation output routing.