CI's JDK8 lane failed: javac resolves API calls against whichever JDK
actually runs it, not the -source/-target level, so compiling on JDK21
(now required to launch sbt 2) bound to JDK9+-only covariant overloads
like ByteBuffer.flip(): ByteBuffer that don't exist on a real JDK8 at
runtime (NoSuchMethodError). This is exactly what the pre-migration setup
avoided by compiling and running each CI lane on the same single JDK.

Restore that by scoping the existing TEST_JAVA_HOME-driven javaHome
setting to the whole build instead of just Test, so sbt forks an external
javac from that JDK for compilation too. Considered --release 8 instead,
but that also hides the sun.misc.Unsafe/sun.nio.ch.DirectBuffer internals
this code intentionally relies on, requiring extra flags for a more
fragile setup than just compiling with the real target JDK.

Wire release.yml/snapshot.yml the same way so published bytecode is
compiled with a real JDK8/JDK11 again, matching their original intent.

The previous fix forked javac onto the target JDK, but missed that the
test suite is Scala, and the Scala compiler always runs in-process inside
the sbt JVM (it doesn't fork the way javac does). With sbt 2 forcing that
JVM to be JDK17+, test code calling e.g. ByteBuffer.flip() directly (as
MessageUnpackerTest.scala does) was still resolving to JDK9's covariant
ByteBuffer.flip():ByteBuffer override, which doesn't exist on real JDK8 ->
NoSuchMethodError at runtime, regardless of the javac fork fix.

scalac's -release flag avoids this the same way --release does for javac,
without needing an ignore-symbol-file escape hatch, since the test sources
never touch JDK-internal APIs the way the main sources' sun.misc.Unsafe
usage does. Verified via javap that the compiled call site now resolves to
Buffer.flip():Buffer instead of the covariant override.

saveToTmpFile called File.createTempFile("testbuf", ".dat", new
File("target")) before tmp.getParentFile.mkdirs() — createTempFile
requires the parent directory to already exist, so under occasional
filesystem timing in CI sandboxes this threw "No such file or directory"
before the directory got created. Surfaced while validating the sbt 2
migration's CI matrix; unrelated to sbt 2 itself, but was intermittently
failing the JDK8 test lane via fail-fast cancellation of the rest of the
matrix.