Kill the process tree and keep the output on execution timeout

[slachiewicz]

On timeout the forked executor now destroys the descendants of the started process before the process itself, and throws ExecutorTimeoutException, a subclass of ExecutorException that carries the last 64 KiB of the grabbed STDOUT and STDERR (stdOutTail(), stdErrTail()). The message keeps its Process timeout: prefix.

Only a tail, because a hung build may have logged far more than should be copied a second time and kept alive with an exception; it is read from the buffer without copying the rest. The grab buffer itself stays unbounded, as before; a caller that needs all of the output passes its own stdOut()/stdErr().

ProcessHandle is called by reflection rather than from src/main/java9: tests and other exploded class paths load from target/classes, where multi-release classes are never used. On Java 8 only the process itself is destroyed, as before.

Fixes #48
Fixes #58

Verified: with process.destroyForcibly() restored, timeoutDestroysTheProcessTree fails (the grandchild's heartbeat file grows from 92 to 110 bytes after the exception); with the whole buffer kept, timeoutKeepsOnlyTheTailOfLargeOutput fails with a 10,488,916-character tail.

Not run on Windows, where the started process is the cmd.exe running mvn.cmd; CI covers it.