perf: optimize levenshtein function

[kazantsev-maksim]

Which issue does this PR close?

• N/A

Rationale for this change

Optimize existing expression.

What changes are included in this PR?

• Introduced a thread_local! scratch buffer pool (LEVENSHTEIN_SCRATCH) that reuses pre-allocated vectors across row evaluations using .resize(). This completely eliminates heap allocations for distance matrices inside the processing loop while remaining strictly thread-safe.
• Added an ASCII Fast-Path (s.is_ascii() && t.is_ascii()) to both levenshtein_distance and levenshtein_distance_with_threshold. When inputs consist entirely of ASCII characters, the functions operate directly on raw UTF-8 byte slices (&[u8]) with zero heap allocations, falling back to the full Unicode char path only when non-ASCII characters are present.

How are these changes tested?

Existing tests.

Benchmark (criterion):