compact into a CSR

[meyerzinn]

LS_CSR compaction is already slow. We might as well re-build the CSR into an actual, bona-fide CSR.

[meyerzinn]

A little bit more on my thinking here: Our ingestion/compaction right now is slower than rebuilding the CSR from an adjacency vector. That's wild and means the contention on the atomic variable is really hurting our ability to do copies in parallel. This PR is a bet that it pays off to just CSR-ify the graph now and get the optimization benefits for workloads like triangle counting.

If we had more time, I would also change the ingestion logic to take an entire batch at once, and regularize it into a CSR to just directly append on the log. Unfortunately I don't think we have time to retool that.

[meyerzinn]

This PR:

• Changes LargeVector to have the same construction semantics as LargeArray (i.e. constructors and destructors are not called by the vector -- it is up to the user to do so if the constructor is nontrivial).
• Removes some extraneous methods from LargeVector
• Adds a new method, addBatchTopologyOnly, which ingests a batched update to an LS_CSR. We basically turn the batched update into a mini-CSR to make copies onto the log more parallel. addEdgesTopologyOnly still exists and continues to work, but we are going to start benchmarking addBatch instead.