`sorensen_dice` byte-length vs character-count mismatch

[meng-xu-cs]

The current sorensen_dice implementation removes whitespace, builds bigrams in character space via s.chars().zip(s.chars().skip(1)), but then uses a.len() / b.len() for the short-length guard and denominator. In Rust, String::len() is a UTF-8 byte count, not a character count, so multibyte characters inflate the denominator and can deflate the score for non-ASCII input.

For example, with a = "café" and b = "cafë":

• Character bigrams for a: (c,a), (a,f), (f,é)
• Character bigrams for b: (c,a), (a,f), (f,ë)
• Shared bigrams: 2
• Correct Sørensen-Dice score: 2 * 2 / ((4 - 1) + (4 - 1)) = 4 / 6 = 0.6666666667
• Current implementation uses byte lengths (5 and 5), so it computes 4 / (5 + 5 - 2) = 4 / 8 = 0.5

So the crate under-reports similarity for multibyte UTF-8 strings.

Incidentally, PR #64: avoid string copy in sorensen dice, fixed it although its title emphasizes performance rather than Unicode correctness.

That PR's diff replaces the byte-based checks and denominator with character-count-based len1 / len2 values, including changing:

• if a.len() < 2 || b.len() < 2 → if len1 < 2 { ... } and if len2 < 2 { ... }
• (a.len() + b.len() - 2) → (len1 + len2 - 2)

However, this PR is in draft mode still.