After the recent changes we're running 5 tests already, some of them
using several VMs. This should cover scheduling to different conductors
well enough, the nova test just adds random failures on top.
This allows reducing the number of test VMs to 3 per testing node
(6 totally), reducing the resource pressure and allowing giving
each VM a bit more RAM.
Also adding missing VM_SPECS_DISK to the subnode configuration.
(cherry picked from commit