Add Armenian language evaluation suite (ArmBench-LLM)

[bugdaryan]

What does this PR do?

This PR adds a comprehensive Armenian (hy) language evaluation suite to LightEval,
covering 24 tasks across classification, extraction, QA, reading comprehension,
summarization, translation, and knowledge/reasoning. All tasks are backed by the
public Metric-AI/ArmBench-LLM-data
dataset on the Hugging Face Hub and are registered under the armenian:* namespace.

Armenian is a low-resource language with almost no standardized LLM evaluation
coverage. This suite gives the community a reproducible way to benchmark models on
native Armenian tasks, with prompts and gold labels written in Armenian.

Context / related reading

The methodology, dataset, and results behind this suite are described in our blog post:
ArmBench-LLM 1.0: Benchmarking LLMs on Armenian Language Tasks.

Tasks added

All tasks are exposed as armenian:<task>:

Category	Tasks
Text classification	topic-14class (SIB-200 topics), sentiment
Orthography / correction	space_fix, punctuation
Named entity recognition	finer, pioner
POS tagging	pos (UD Armenian-ArmTDP)
Open / in-context QA	arak, ms_marco, squad
Reading comprehension & MCQA	belebele, scientific, syndarin, dream, include, hartak
Summarization	email, conversation
Paraphrasing	paraphrase
Machine translation	short_sentences_translation
Knowledge & reasoning	mmlu_pro, exam_history, exam_literature, exam_math

Metrics added

Armenian tasks need scoring robust to Armenian script and answer formatting, so this
PR introduces:

• Armenian-specific metrics (metrics/armenian_metrics.py): exam accuracy, MCQA
  accuracy, MMLU-Pro accuracy, NER span-overlap F1, POS accuracy, and an
  Armenian-tuned BERTScore.
• Text-tagging metric (TextTaggingMetric + text_tagging_metric): character
  n-gram F1 matching for keyword/keyphrase extraction tasks.
• Extraction helpers (metrics/utils/armenian_eval_utils.py): regex-based answer
  extractors for letter/numeric multiple-choice and free-form numeric answers.

Other changes

• transformers_model.py: wrap batch[0].stop_sequences in list() before
  concatenating with the EOS token, so generation works when stop_sequences is a
  tuple rather than a list.

How to run

lighteval accelerate \
  "model_name=<your-model>" \
  "lighteval|armenian:mmlu_pro|0|0"