ByteStrings: cross-boundary swar

[pjfanning]

The ByteString2 class (2-fragment ByteString) had cross-boundary read methods that used 8 individual byteAt calls for Long reads. Each byteAt call has a branch (check which fragment). The optimization:

• Long reads (BE/LE): Build two sub-long values from byteAtUnchecked calls (which skip the bounds check branch) and combine them with bitwise OR. This reduces the branch overhead from 8 per Long to 8 virtual calls (one per byte, no extra branching per call) plus pure arithmetic assembly.
• Short/Int reads: Kept using byteAt via SWARUtil helper methods. With only 2-4 bytes needed, the branch overhead is minimal and not worth adding complexity.

Why not direct array access: The first/second fields are typed as ByteString (trait), and .bytes/.startIndex are private[pekko] on the concrete subclasses but not accessible from the nested ByteString2 class context. byteAtUnchecked provides the same bounds-check-skipping benefit through virtual dispatch.