Unexpected behavior with default compilation set

[bachase]

Describe the bug
While reviewing unitaryfund/ucc-demo#1, @jordandsullivan noticed that a simple Bell state circuit was transforming the circuit even though it was already in the target gate basis:

Debugging a bit, this appears to be due to forcing consolidation of blocks, after which unitary synthesis expands to the different gateset:

ucc/ucc/transpilers/ucc_defaults.py

Line 73 in d92c341

self.pass_manager.append(ConsolidateBlocks(force_consolidate=True))

To Reproduce
Steps to reproduce the behavior:

1. Run the circuit in the unit tests

   ucc/ucc/tests/test_compile.py

   Lines 19 to 22 in d92c341

   	circuit = QiskitCircuit(2)
   	circuit.h(0)
   	circuit.cx(0, 1)
   	result_circuit = compile(circuit, return_format="original")

To debug, I edited UCCDefault1.run to print the circuit after each pass:

def callback_func(**kwargs):
    from qiskit.converters import dag_to_circuit

    pass_ = kwargs["pass_"]
    dag = kwargs["dag"]
    print(f"Pass: {pass_.name()}")
    print(dag_to_circuit(dag))
    print("-" * 80)

class UCCDefault1:
....
   def run(self, circuits):
       return self.pass_manager.run(circuits, callback=callback_func)

Expected behavior
We would expect the circuit to stay the same.

[bachase]

@jordandsullivan and @Misty-W , removing the force_consolidate gives the expected result, but I'm wondering what the original reason was behind setting that to True in the first place. Do either of you know? I'll make a branch with that change and run the benchmarks to see the impact.

On the flip side, there's an interesting question of why the compiler didn't re-optimize back to the simpler circuit after that pass, versus the more complex one it ended up with.

But stepping back, for the next milestone, it might be worth revisiting these choice of default passes, and ensuring we capture somewhere the reasons, or at least types of circuits, where they perform well.

[bachase]

ef0e5c9 has the benchmark results.

Pulling out the gate counts -

compiler	circuit_name	raw_multiq_gates	ucc baseline(force consolidate)	ucc_dont_force_consolidate	qiskit-default
ucc	qaoa_barabasi_albert	1176	1176	1176	1176
ucc	qft	10050	2740	3244	3244
ucc	qv	15000	14856	14856	14856
ucc	square_heisenberg	2160	540	540	540
ucc	prep_select	9744	9702	9710	9708
ucc	qcnn	388	388	388	388

So qft gate count is definitely worse without the consolidation. I added qiskit-default as a baseline, as it shows that ucc without the consolidation is the same gate count reduction for qft as qiskit (ignoring compilation time for now).