Eliminate duplicate chunk downloads using recipes

[bgilbert]

Allow packages to contain "recipe" metadata giving hashes of memory and disk image chunks. Use the recipe on the client to avoid redundant chunk downloads, e.g. by finding an identical chunk cached from another VM. Demand-fetch recipe chunks to avoid excessive up-front download cost.

For this to be useful, there must be low overhead and significant intra- or inter-VM redundancy. Consider selectively enabling recipe processing based on bandwidth and round-trip latency to the server. Investigate the amount of redundancy remaining after xz memory image compression (probably none) and qcow2 disk image compression.