fix: handle non-ASCII text in ROUGE-1 eval scorer

[guptaishaan]

The rouge_score library's default tokenizer matches only ASCII alphanumeric
characters ([a-z0-9]+) and silently discards every other character. As a
result, text written in non-Latin scripts such as Thai (สวัสดี), Chinese,
Arabic, or any other language that uses non-ASCII code points is tokenized to
an empty list, which causes _calculate_rouge_1_scores to return a score of
0.0 even when the candidate and reference strings are identical.

The fix introduces a small _UnicodeTokenizer class that extends
rouge_score.tokenizers.Tokenizer. Its tokenize method uses the regex
[a-z0-9]+|[^\x00-\x7f]+ to preserve contiguous runs of non-ASCII characters
as individual tokens while keeping the same ASCII-alphanumeric filtering for
Latin text. The _calculate_rouge_1_scores helper is updated to pass this
tokenizer to RougeScorer instead of using use_stemmer=True with the
default ASCII-only tokenizer. Two regression tests for identical and
non-matching Thai strings are added to the existing test module.

Fixes #3111