Use custom JVM versions during Scala compilation

[lihaoyi]

Follow up from #3480

Currently JavaModule#compile with a custom JVM relies on Zinc forking a JVM for each Java compilation, which is slow, and Zinc does not support forking JVMs for ScalaModule#compile and it would anyway be prohibitively slow (upstream discussion sbt/zinc#1498).

The correct solution for this is to move ZincWorkerImpl into a long-lived worker subprocess when def javaHome is overridden. This will allow us to keep the Java compilation process warm so it can benefit from the fast in-memory compiles, while also supporting Scala compilation on custom JVM versions

This will likely require an overhaul of the ZincWorkerApi.scala to dis-entangle things sufficiently that we can impose a subprocess boundary between the caller and callee. This would break binary compatibility and would need to wait for 0.13.0

[roman-mibex-2]

The mentioned relevant Zinc issue: sbt/zinc#1498