Add nova cache layer for different runs to reduce workload if necessary

The currently behavior for compute polling task is to call nova_cli to get all instance infos. If the instances in nova-compute didn't change, this will be a waste of time and workload to nova/glance client.

During large deployment, say thousands of compute agents, it will be better to provide a configurable nova cache layer to save instances information and flush it periodically, it can dramatically reduce workload from nova/glance client during large scale deployment. Especially for some scene that don't need high SLA and allow some delay, it's more useful.

So here I think it's worthy to implement a configurable nova cache layer.