Remove deepcopy usage from circuit_to_dag()

[mtreinish]

Summary

The circuit_to_dag() function was previously using a deepcopy to make
copies of all input circuits. This could get quite expensive as it has
to update all the references. However, this deepcopy was only being used
to create a fresh copy of the instructions objects that were part of the
circuit. This instructions are passed directly to the dag in the output
so we can't pass by reference otherwise changes to the dag will break
the circuit. To avoid the performance hit with a deepcopy this commit
changes the function to instantiate new instruction objects right before
we add them to the dag.

Details and comments

Fixes #1528

[jaygambetta]

do we have speed improvement numbers

[mtreinish]

I've just done a small synthetic test running this function on a freshly loaded large circuit object (loaded from a 3500 line qasm file) and converting that to a dag. It was ~1.6x as fast for that quick conversion going from ~0.21 secs to ~0.13 secs. The raw data is:

With Patch:
0.128338
0.129884
0.125828
0.125982
0.126041
0.129518
0.126339
0.125583
0.126122
0.127914
0.125035
0.127082
0.126778
0.125588
0.130447

Without Patch:
0.205295
0.206874
0.207197
0.208566
0.214064
0.208503
0.208941
0.207808
0.210187
0.205799
0.207779
0.213378
0.209816
0.211099
0.205749

If you've have any specific benchmarks you'd like me to run I can quickly throw it together.

[mtreinish]

I ran transpile and compile on the same circuit from qasm and ran benchmark runs on those. For transpile() its a ~1.4x speed up going from ~0.27 secs to ~0.19 secs and for compile() it's a ~1.2x speed up going from ~1.33 secs to ~1.12 secs. The raw data from that testing is here: compile_transpile.txt

Heh, not quite as drastic as the >1000x speedup from the last time we did a patch like this in #1431