Previously, there was a race condition between the RLIMIT_CPU timeout
(which resulted in a SIGKILL return code, meaning test failure) and the
subprocess.Popen.communicate timeout (which resulted in a
subprocess.TimeoutExpired exception that was converted to a SIGXCPU
return code, meaning test success). This meant that .inf.java and
.inf.mj tests would semi-consistently fail.

This fixes that issue by increasing the RLIMIT_CPU timeout to five
seconds longer than the subprocess.Popen.communicate timeout. I won't
pretend that this is anything but a quick dirty fix, but at least it
matches the style of the rest of mjtest.