Level 1 chooses between trival and dense layout selection, depending on the layout score

[1ucian0]

On top of #3035, #3018, and #3321 (therefore, on hold).

Fixes #2845

[1ucian0]

@kdk suggested a great possible optimization to rollback_if: Do not rollback (e.i. deepcopy) the dag if all the passes in the pass set are analysis pass (which is the case when layout selection).

Good news! Adding this optimization was trivial (see ec8e31a). I will rerun the tests.

[1ucian0]

(because I'm getting full different set of values (probably related with changes in master or my machine is just less loaded) here are the new outputs:

Without rollback

The result is: 58.22561879

With rollback

The result is: 58.96733729300001

Comparing with level 2

Level 2 is running in: 346.168352814

Comparing against master

1.74x slower for qft.QftTranspileBench.time_ibmq_backend_transpile(2)

[1ucian0]

Closing in favor of #3657