perf: avoid redundant bounds checks on ByteString indexed reads

[pjfanning]

Motivation

ByteString.apply(idx) bounds-checks before delegating to byteAtUnchecked, and for the ByteStrings composite it also resolves the absolute offset to a fragment. Two call sites under src/main pay that check when the caller has already established the bound:

• ByteStringParser.ByteReader.readByte guards with off < input.length and then calls input(off), so the bound is tested twice. Its siblings readShortLE / readIntLE / readLongLE already use the *Unchecked variants after their own guard, so readByte was the odd one out. ByteReader backs gzip/deflate header parsing, TLS and framing.
• TcpDnsClient.decodeLength read the two length-prefix bytes with two apply calls: two bounds checks, and for a composite ByteString two fragment lookups.

These are the only two places under src/main that index a ByteString outside ByteString.scala itself and bench-jmh.

Modification

readByte now calls input.byteAtUnchecked(off). off starts at 0 and only ever moves forward, so the existing guard is sufficient.

decodeLength now uses data.readShortBE(0) & 0xFFFF. readShortBE keeps the bounds check, so short input still throws IndexOutOfBoundsException as before, and it is SWAR-optimised for the single-array ByteString1C / ByteString1 that the common single-chunk read path produces.

Both are internal (private[pekko] / private[internal]); no public API or binary shape change.

Result

One bounds check per byte instead of two in the stream parser, and one checked 16-bit read instead of two indexed reads for the DNS length prefix. Behaviour is unchanged.

Worth noting for reviewers: ByteStrings.apply is already cheap for sequential access thanks to the fragmentHint memoisation, so this is a small constant-factor cleanup rather than a complexity fix. I did not switch ByteReader to an iterator — it needs off as a random-access cursor for take, slice and fromStartToHere.

Tests

• sbt "stream-tests/testOnly org.apache.pekko.stream.io.ByteStringParserSpec org.apache.pekko.stream.scaladsl.CompressionSpec org.apache.pekko.stream.io.compression.*" — 106 succeeded, 0 failed
• sbt "actor-tests/testOnly org.apache.pekko.io.dns.internal.TcpDnsClientSpec" — 8 succeeded, 0 failed
• sbt "actor/mimaReportBinaryIssues" "stream/mimaReportBinaryIssues" — success
• Native scalafmt run on the two changed files.
• No new tests: both changes are behaviour-preserving refactors of internal code, covered by the existing specs above.
• Not benchmarked. The change is a strict reduction in work per read, not an algorithmic change — say the word if you want JMH numbers before merging.

References

None - found by reviewing ByteString indexed access under src/main