Otherwise, deployments with multiple VMs try to write to the same image.

Also, do the temp_image_path calculation only once.

`MachineState.run_command()` passes SSH flags to `self.ssh.run_command()`.
However, `self.get_ssh_flags()` is already registered as a `ssh_flag_fun` in the
class `__init__()` function, so `ssh_util.SSH` already uses it to get the flags
when initiating a connection. This lead to the SSH flags being duplicated, which
causes an error for some flags (e.g. the `-J` flag, which can only be specified once).

Offers better separation, especially when additional features will be added.

This helps in situations when there's no network connectivity to the guest, e.g.
when the hypervisor host can be reached via a VPN, but the guest host cannot.

This mainly adds driver support for virtio drivers to the initial image of the
guest provisioning.

This is a WIP!

- Replaced `deployment.libvirtd.networks` option with a submodule to allow not
  only (libvirt) network names, but other networking types as well.
- Domain XML was adjusted accordingly to incorporate the parameters from the new
  `networks` submodule.
- Added the qemu guest agent to guests to allow for out-of-band
  communication (no need for network connectivity) with the hypervisor.
- Guest IP (for provisioning after guest has started) is no longer determined by
  waiting for the guest to get a DHCP lease in the hypervisor libvirt network.
  If the guest has a static IP, it won't ask for a DHCP lease. Also, for bridged
  networking, we probably will not have access to the DHCP server.
- Instead, the address of the first interface is retrieved from libvirt using
  the `VIR_DOMAIN_INTERFACE_ADDRESSES_SRC_AGENT` method, which can now be done
  because of the newly added qemu guest agent.