Consider deprecating the VirtualGraphComposite

[arcondello]

Reasons:

• It is not (as far as I know) widely used
• It requires significant QPU time to configure - we even have a warning to that effect
• It is less relevant now with the lower noise processors and does not implement the latest protocols
• Removing it would allow us to remove the dependency on dwave-drivers, and remove a lot of caching boilerplate

[JoelPasvolsky]

My 2c for this debate: I think that :func:~dwave.embedding.chain_strength.uniform_torque_compensation adds urgency, and that we need to either deprecate or update VG's chain_strength algo because my guess (based on just a small sample) is that despite improvements from adjusting flux biases, results from the max negative chain strength do not compare favorably with the calculated chain-strength from uniform_torque_compensation .

A 50-node clique BQM with biases generated by dimod.generators.randint(50, 'SPIN', low=0, high=5) and scaled down to the J range, reusing the embedding with VG produced unsatisfying results:

1. Chimera Non-VG: VG:

2. Pegasus Non-VG: VG: