Fix helm_normalizer mis-scoring numbers (homogeneize before remove_punc)

[iamsharduld]

What

In helm_normalizer (the quasi-exact-match / HELM normalizer), remove_punc runs before
homogeneize_numbers, so it strips the decimal point before the float() cast. The result is wrong
on numeric answers:

from lighteval.metrics.normalizations import helm_normalizer
from lighteval.metrics.metrics_sample import ExactMatches
em = ExactMatches(normalize_gold=helm_normalizer, normalize_pred=helm_normalizer)

em.compute_one_item("10", "1.0")   # 1  -> distinct numbers scored as EXACT MATCH (false positive)
em.compute_one_item("3.14", "314") # 1  -> false positive
em.compute_one_item("1.0", "1")    # 0  -> false NEGATIVE, contradicting homogeneize_numbers' docstring

Root cause: '1.0' → remove_punc → '10' → float → '10.0', and '3.14' → '314' → '314.0',
so different numbers normalize to the same string while 1.0 and 1 normalize differently.

Fix

Run homogeneize_numbers before remove_punc, so float() sees the original token. Distinct
numbers then stay distinct and equal-but-differently-formatted numbers match, per the docstring intent.
Non-numeric tokens are unaffected (homogeneize_numbers returns them unchanged either way).

Tests

tests/test_unit_base_metrics.py::test_quasi_exact_match_numbers — 1.0 == 1 matches; 10 != 1.0
and 3.14 != 314 don't. Fails before, passes after; the existing test_quasi_exact_match (sentence
text) still passes.