Parameter binding substantially slower on wrapped circuit vs unwrapped

[kdk]

What should we add?

Calling QuantumCircuit.bind_parameters can have drastically different performance depending on whether or not the circuit contains wrapped instructions. For example, comparing the time to build and bind parameters for an EfficientSU2 circuit, where in the first case, we call bind_parameters directly on EfficientSU2, and in the second, after unwrapping EfficientSU2 via a call to decompose:

from simple_benchmark import benchmark

def _build_wrapped(num_layers):
    qc = EfficientSU2(100, entanglement='linear', reps=int(num_layers))
    qc.bind_parameters({p: math.pi/2 for p in qc.parameters})
    
def _build_unwrapped(num_layers):
    qc = EfficientSU2(100, entanglement='linear', reps=int(num_layers)).decompose()
    qc.bind_parameters({p: math.pi/2 for p in qc.parameters})
    
funcs = [_build_wrapped, _build_unwrapped]
b = benchmark([_build_wrapped, _build_unwrapped],
              {i: i for i in np.linspace(1, 100, 5)},
              'num_layers',
              function_aliases={_build_wrapped: 'wrapped', _build_unwrapped: 'decomposed'}
             )

Related to #10269 . During profiling @mtreinish found most of the time for the wrapped case spent in recursive calls to https://github.com/Qiskit/qiskit-terra/blob/main/qiskit/circuit/quantumcircuit.py#L2933-L2944 .