This commit adds a cache for parameter free gate decompositions in
the Unroll3qOrMore transpiler pass. If there are repeated gates in a
circuit that don't have any parameters, then when we decompose the
circuit a single time that decomposition will be valid for all instances
of that gate. This doesn't hold true for parameterized gates because the
parameter value will determine the decomposition used. This greatly
improves the runtime performance of the pass when there are repeated >=3
qubit gates in the circuit.

In a future revision we can investigate adding the parameter values to
the cache lookup. But the hit rate would likely be fairly low in such
cases because it would only provide runtime improvements if there are
repeated >3 qubit gates with the same exact parameter values, which
seems fairly rare in practice.

This is based on top of Qiskit#10702 and will need to be rebased after Qiskit#10702
merges.

This commit adds a cache for parameter free gate decompositions in
the Unroll3qOrMore transpiler pass. If there are repeated gates in a
circuit that don't have any parameters, then when we decompose the
circuit a single time that decomposition will be valid for all instances
of that gate. This doesn't hold true for parameterized gates because the
parameter value will determine the decomposition used. This greatly
improves the runtime performance of the pass when there are repeated >=3
qubit gates in the circuit.

In a future revision we can investigate adding the parameter values to
the cache lookup. But the hit rate would likely be fairly low in such
cases because it would only provide runtime improvements if there are
repeated >3 qubit gates with the same exact parameter values, which
seems fairly rare in practice.

This is based on top of Qiskit#10702 and will need to be rebased after Qiskit#10702
merges.

This commit adds a cache for parameter free gate decompositions in
the Unroll3qOrMore transpiler pass. If there are repeated gates in a
circuit that don't have any parameters, then when we decompose the
circuit a single time that decomposition will be valid for all instances
of that gate. This doesn't hold true for parameterized gates because the
parameter value will determine the decomposition used. This greatly
improves the runtime performance of the pass when there are repeated >=3
qubit gates in the circuit.

In a future revision we can investigate adding the parameter values to
the cache lookup. But the hit rate would likely be fairly low in such
cases because it would only provide runtime improvements if there are
repeated >3 qubit gates with the same exact parameter values, which
seems fairly rare in practice.

This is based on top of Qiskit#10702 and will need to be rebased after Qiskit#10702
merges.

In the Unroll3qOrMore pass internally is quite simple it iterates over
every operation in the circuit and for any operation that uses >= 3
qubits and recursively decompsing it into all 1 and 2 qubit operations,
converting it to a DAGCircuit and substituting the >=3 qubit gates with
the equivalent circuit. However, during this DAGCircuit conversion step
we're spending a large amount deep copying the operation objects.
However, we don't need to do this because nothing in the circuit will be
reused with shared references so we can skip the copying and just pass
the operation objects by reference onto the DAG (as that's all we need).
This commit makes that change by using the `copy_operations` flag we
introduced in Qiskit#9848 on circuit_to_dag() to disable the internal copying.