Fix asdiv prompt truncating multi-character gold answers

[vineethsaivs]

Problem

asdiv_prompt sets choices to a bare string instead of a list:

def asdiv_prompt(line, task_name: str = None):
    return Doc(
        task_name=task_name,
        query=f"{line['body']}\nQuestion:{line['question']}\nAnswer:",
        choices=line["answer"].split(" (")[0],   # bare string, e.g. "35"
        gold_index=[0],
    )

Doc.choices is a list[str], and Doc.get_golds() indexes it with the gold index:

for gold_ix in as_list(self.gold_index):
    golds.extend(as_list(self.choices[gold_ix]))

When choices is the string "35", choices[0] indexes the string and yields "3", so the gold answer is silently truncated to its first character. ASDiv answers have the form "<value> (<unit>)" (which is why the .split(" (")[0] is there), so the extracted value is usually multi-character:

answer "35 (apples)" -> gold ["3"]   (should be ["35"])
answer "128"         -> gold ["1"]   (should be ["128"])
answer "3.5 (m)"     -> gold ["3"]   (should be ["3.5"])

get_golds() feeds the reference into the task's exact_match metric, so a model that correctly outputs 35 is scored against gold 3 and marked wrong. Every sibling generative task wraps the value in a list (unscramble, numeracy, arithmetic, ...); asdiv is the only one missing the brackets.

Fix

Wrap the extracted answer in a list, matching the sibling tasks and the Doc.choices: list[str] contract:

choices=[line["answer"].split(" (")[0]],

Test

Adds tests/unit/tasks/test_asdiv.py, which calls asdiv_prompt on a multi-character answer and asserts doc.choices == ["35"] and doc.get_golds() == ["35"]. It fails before (get_golds() returns ["3"]) and passes after.