compute agent interaction with hypervisor layer independent of nova internals

Currently the compute agent uses the nova hypervisor driver code in order to poll the virt infrastructure.

This has led to a very messy dependency on nova internals, which causes frequent breakage on trunk.

Since we use only a tiny subset of the nova virt code, we should instead code directly to the underlying hypervisor layer (e.g. the libvirt-python binding) as opposed to relying on nova internals which are subject to rapid change.

This mechanism should be pluggable so that a similar approach could be used to interact with xenapi (so as to free us from the libvirt-only rut).