Duplicate gates with higher-order gate functions in structured programs

The normal mechanism in PennyLane that "un-queues" a gate when supplied to a HOF like qml.adjoint does not work when the original and modified gate appear in different program scopes, such as across control flow segments. The reason is that Catalyst using different tapes for each section.

import pennylane as qml

@qml.qjit
@qml.qnode(qml.device("lightning.qubit", wires=1))
def circuit(b: bool):
    g = qml.H(0)

    @qml.cond(b)
    def apply():
        qml.adjoint(g)
    apply()

    return qml.state()

print(circuit(1))  # should not read [1, 0]

Start by reproducing the example with qml.qjit, qml.cond, and qml.adjoint, then trace how Catalyst handles gates captured in separate program scopes or tapes. Compare the resulting state with the expected behavior and add coverage for the cross-control-flow case; done means the original and modified gates are not duplicated.