Remove deepcopy usage in dag_to_circuit

[mtreinish]

Summary

This commit attempts to improve the performance of dag_to_circuit by
removing the deepcopy usage. Instead this switches to creating a
circuit object using the normal constructors while iterating over the
dag.

Details and comments

[nonhermitian]

When done, can you do a timeit to see the improvement?

[mtreinish]

Yeah, once I get this passing all the tests I'm planning to add a comment here with some benchmarks.

But, at least on this version of the PR I've done some preliminary timing with compiling a large qasm file (with like ~3.5k lines) from aqua that @ajavadia provided and got a significant speedup in that case. Calling compile() took ~3.1secs with this patch and ~3310.3 secs without this patch.

[jaygambetta]

Can you do the timing on transpile not complie

[mtreinish]

Ok, I've fixed the issue caused by cx_base vs cx now the only failing test is: https://travis-ci.org/Qiskit/qiskit-terra/jobs/464115683#L1397 Not sure what I did to break it there, but the expected counts are nowhere near what they should be.

[mtreinish]

I'm still running numbers (without this patch transpile can take a long time) but so far I've finished the tests with and without this patch just parsing the same large qasm file from aqua I mentioned before and getting an unrolled dag from that and then timing how long dag_to_circuit() takes on that dag. With this patch that operation ~0.40 secs and without it took ~0.95 secs. The raw numbers for that test were:

With Patch:
0.397959
0.411044
0.405604
0.410246
0.401071
0.406355
0.404786
0.406011
0.403361
0.409036
0.405321
0.416254
0.400951
0.404211
0.400891

Without Patch:
0.950606
0.961483
0.988895
0.957507
0.952986
0.955229
0.952623
0.963022
0.959569
0.954256
0.95801
0.959481
0.954618
0.95338
0.952434

I'm running the transpile benchmarks using that same large qasm file now. I've finished the runs with this patch and it takes ~1.4 secs to run transpile(). The runs without this patch are taking a lot longer, the script is still running (only doing 5 samples) but it's taking ~3300 seconds to run transpile() each time so far.