Fewer allocations in Transliterator runtime

[skius]

Segments (regex capture groups) currently use a Vec<String>. That should almost certainly be a SmallVec, it could potentially even be a ShortSlice.

Component-based hardcoded transliterators like NFC allocate, because the Writable/Write based API is incompatible with the in-place modifications we do when transliterating. I don't know how exactly the normalizing algorithms work, but if it's forward-only (no lookback), we could expose that and avoid allocating an intermediate String.