Precompute clique decomposition

For larger systems, there are two overheads in building an `OPFModel{SparseSDPOPF}`:
1. Computing the clique decomposition when building the `OPFData` (repeated at the beginning of each sampling job).
2. Linking the variables representing the same bus pair (repeated for each seed).

For 2869_pegase, `build_opf` took about 10 minutes on my laptop, where Step 1 above took relatively more time. They could become a large overhead if the allocated time is relatively short.

Solutions:
1. We can precompute and save the cliques for some of the PGLib systems. Computing the cliques offline allows us to use more sophisticated algorithms.
2. Variables linking mappings can also be pre-computed.

(What if one were to work with one of these systems, but with a different branch status? The clique decomposition of the original network would still be valid; it would only potentially include some redundant edges.)