fix(evals): Exact match float int mismatch

[andrei-rusu]

No description provided.

[cristipufu]

This fix doesn't cover the case where target_output_key is "*" (the default).

When target_output_key is "*", _get_actual_output / _get_expected_output return the full dict, so the flow becomes:

actual_output = str({"result": 27.0})    # "{'result': 27.0}"
expected_output = str({"result": 27})     # "{'result': 27}"

float("{'result': 27.0}")  # ValueError → falls back to string comparison → mismatch

The float() try/except only helps when a specific key is extracted (scalar value). With "*" the whole dict is stringified and float() always fails, so it's the same broken path as before.

Repro: any evaluator JSON with "targetOutputKey": "*" comparing {"result": 27} expected vs {"result": 27.0} actual → scores 0%.

The fix probably needs deep structural comparison with numeric normalization for the dict case, e.g. recursively normalizing int/float before comparing.