fix(afrobench): pass the document to the XL-Sum prompts

[alleyxd]

All 36 AfroBench XL-Sum tasks (xlsum_<lang>_prompt_{1,2,3}, 12 languages) ask the model to summarize the word text instead of the article. The generator in afrobench/xlsum/utils.py builds the prompts with r"{{'text'}}\n": the quotes make Jinja render the string literal 'text' rather than the text field, and the raw string keeps \n as two characters.

Rendering xlsum_swahili_prompt_2 with apply_template and a doc whose text is THE ARTICLE BODY, at d6de816:

'Summarize the document below in triple backticks and return only the summary and nothing else.\n```text```\\n'

After:

'Summarize the document below in triple backticks and return only the summary and nothing else.\n```THE ARTICLE BODY```\n'

The fix is in the generator (all five prompt templates, including the two that have no YAMLs checked in); the 36 YAMLs were regenerated with python utils.py --mode prompt_1 prompt_2 prompt_3, so only doc_to_text changes in each. Since every score from these tasks changes, the three prompt_*/xlsum bases go from version 1.0 to 2.0, with a changelog line in the README.

Verification

$ pytest tests/test_afrobench_xlsum.py -q
37 passed
# fix reverted, test kept: 36 failed, 1 passed

$ pytest tests/test_tasks.py -q -k "afrobench or xlsum"
37 passed, 16 deselected

pre-commit run passes on the changed files.

[CLAassistant]

All committers have signed the CLA.