fix(json): do not treat booleans as numbers in JSONDiff

[fei (feiiiiii5)]

Fixes the case where a JSON boolean compared against an integer scores a perfect match.

JSONDiff picks the numeric scorer with a bare isinstance check:

elif (isinstance(o1, int) or isinstance(o1, float)) and (isinstance(o2, int) or isinstance(o2, float)):
    return self.number_scorer.eval(o1, o2).score

bool is a subclass of int, so booleans take that branch. NumericDiff(True, 1) is 1 - |1 - 1| / (1 + 1) = 1.0, and the same happens for 0 vs False:

JSONDiff().eval({"flag": True}, {"flag": 1}).score    # 1.0
JSONDiff().eval({"flag": False}, {"flag": 0}).score   # 1.0
JSONDiff().eval([True], [1]).score                     # 1.0

For a scorer, a boolean and an integer are different JSON values, and a model that returned true where 1 was expected is being marked fully correct.

The TypeScript implementation tests typeof v === "number", which is false for booleans, so it never entered that branch and scored the same input 0.0. AGENTS.md says the two implementations share the same scorer behavior; here they did not.

Excluding bool from the numeric check sends Python booleans to the same JSON.stringify fallback the TypeScript version already reaches. Running the same inputs through both published autoevals 0.3.0 packages, they now agree:

input	before (py)	before (ts)	after
{"flag": true} vs {"flag": 1}	1.0	0.0	0.0
{"flag": false} vs {"flag": 0}	1.0	0.0	0.0
[true] vs [1]	1.0	0.0	0.0
{"flag": true} vs {"flag": true}	1.0	1.0	1.0
{"flag": "true"} vs {"flag": true}	0.6667	0.6667	0.6667

The last two rows matter: equal booleans still score 1.0, and the string "true" still does not match the boolean.

Test: pytest py/autoevals/test_json.py -k booleans fails on main at 9546b28 with AssertionError: {'flag': True} scored equal to {'flag': 1} and passes on this branch. pytest py/autoevals/test_json.py is 6 passed. The other non-LLM test files pass; test_partial fails on main as well, from a missing OPENAI_API_KEY in this environment. black, ruff and codespell are clean on both changed files at the versions in .pre-commit-config.yaml. I did not run the LLM-backed tests, which need credentials.

The list branch divides by max(len(o1), len(o2)) while iterating zip(o1, o2), so extra elements lower the score rather than being ignored. That is intended and is identical in both implementations, so I left it alone.