common: fix match_case misidentifying all-caps contractions as capitalized

[nac7]

Fixes #4313.

Problem

match_case in src/helm/common/general.py iterates over all characters when checking islower()/isupper(), including apostrophes. In Python, "'".islower() and "'".isupper() both return False (apostrophes have no case). This breaks the all-caps detection for any word containing an apostrophe:

# Before fix:
match_case("AIN'T", "is not")   # → "Is not"  ✗ (should be "IS NOT")
match_case("DON'T", "do not")   # → "Do not"  ✗ (should be "DO NOT")
match_case("CAN'T", "cannot")   # → "Cannot"  ✗ (should be "CANNOT")

For "AIN'T": all(ch.isupper() for ch in "AIN'T") is False because "'".isupper() is False, so the all-caps branch is skipped. The code falls into the capitalized branch (source_word[0].isupper() is True), producing "Is not" instead of "IS NOT".

This affects ContractionPerturbation, ExpansionPerturbation, and SynonymPerturbation, which all call match_case to preserve source-word casing. All-caps contractions in benchmark inputs produce wrong-case augmented outputs.

Fix

Filter to only alphabetic characters before case-checking. Non-alpha characters (apostrophes, digits, punctuation) are ignored when determining the case category of a word:

alpha_chars = [ch for ch in source_word if ch.isalpha()]
if not alpha_chars:
    return target_word
if all(ch.islower() for ch in alpha_chars):
    return target_word.lower()
if all(ch.isupper() for ch in alpha_chars):
    return target_word.upper()
if alpha_chars[0].isupper():
    return target_word.capitalize()
return target_word

Tests

Added to test_perturbation.py:

• test_match_case_all_caps_with_apostrophe — "AIN'T" / "DON'T" / "CAN'T" → "IS NOT" / "DO NOT" / "CANNOT" (regression cases)
• test_match_case_lowercase_with_apostrophe — lowercase contractions still produce lowercase
• test_match_case_capitalized_with_apostrophe — capitalized contractions still capitalize
• test_match_case_no_alpha_chars — empty / digit-only source → target unchanged
• test_match_case_regular_words — regular words (no apostrophe) still work correctly