Added the creation of a local TF service to handle local CPU device (to allow running XRT on local workstation).
The TF DIRECT_SESSION (the ones used so far) isolates the different sessions memory domains, and hence the int64 memory handles handed over by one session, are not valid in the other.
This is no issue when using GRPC sessions, as the memory context is unified.
Since we need multi-session (as Run() is not thread safe), we create a local TF/CPU service if the user configured as such.