In SuperWord, there are some heuristics that predict the profitability of vectorization. These are not always correct, and we may want to benchmark them thoroughly and improve them in the future.

To facilitate this benchmarking, I propose we add a diagnostic flag:
AutoVectorizationOverrideProfitability
0: never profitable, i.e. abort SuperWord without vectorization
1: Use profitability heuristics to determine profitability
2: ignore profitability heuristic, vectorize anyway (if all correctness checks pass, of course)

This flag will also help us when we are working on the cost model.

And: in some cases, we end up not vectorizing because it is not possible. But the unrolling mechanism behaves different with UseSuperWord enabled and disabled. Sometimes the performance with it enabled is worse than without, because of over-unrolling.