fix: correct element-wise comparison in words_diff_ratio

[lxcxjxhx]

Summary

Fix words_diff_ratio to return correct element-wise comparison ratio instead of always returning 1/num_words.

Problem

The method words_diff_ratio in AttackedText compares two word lists using self.words != x.words. Since self.words is a Python list, this comparison returns a single boolean value (True/False) rather than an element-wise comparison array.

# Before (incorrect)
return float(np.sum(self.words != x.words)) / self.num_words
# self.words != x.words returns True (single boolean)
# np.sum(True) = 1
# Result: always 1/num_words, regardless of actual differences

Fix

Convert lists to numpy arrays before comparison to enable element-wise comparison:

# After (correct)
return float(np.sum(np.array(self.words) != np.array(x.words))) / self.num_words

Testing

from textattack.shared.attacked_text import AttackedText

text1 = AttackedText({"text": "the quick brown fox"})
text2 = AttackedText({"text": "the slow brown fox"})

# Before fix: 0.25 (1/4, incorrect)
# After fix: 0.25 (1/4, correct - only "quick" vs "slow" differs)

text3 = AttackedText({"text": "a b c d"})
text4 = AttackedText({"text": "w x y z"})

# Before fix: 0.25 (1/4, incorrect)
# After fix: 1.0 (4/4, correct - all words differ)

Fixes #787