Improve circuit compilation performance

[ohuettenhofer]

This PR speeds up working with and compiling large circuits.

Here is a simple benchmark to test this:

Code

import tequila as tq
import random
from time import time

if __name__ == "__main__":
    random.seed(0)
    start = time()

    circuit = tq.QCircuit()
    for i in range(100):
        subcircuit = tq.QCircuit()
        for j in range(100):
            qubits = random.sample(range(10), 3)
            subcircuit += tq.gates.Toffoli(qubits[0], qubits[1], qubits[2])
            subcircuit += tq.gates.Ry(0.5, random.choice(range(10)))
        circuit += subcircuit

    daggered = circuit.dagger()
    circuit += daggered

    tq.simulate(circuit, backend="qulacs", optimize_circuit=False)

    print(f"Time: {time() - start:.3f} s")

Having optimize_circuit=False is necessary, without it the Qulacs optimization will take very long.

Before this pull request, the script took about 17 seconds on my machine.

The first change is to revert #360, because I didn't realize how big the performance impact of the deepcopy call here is.
I don't like removing this, but it reduces the runtime to around 10 seconds, and it seems like it was fine without the copy for years before I made that pull request.
I extended the warning message to explain the specific case that made me add the copy.

The second change is optimizing the replace_gates function, because it is currently implemented in an inefficient way that performs a lot of unnecessary copying. This change reduces the runtime of the script to about 3 seconds.

I also fixed a problem in the QubitWaveFunction length method where it also counted zeros, and added a function for BitString that returns the stored integer in a specified bit ordering.

[kottmanj]

Thanks!
Looks like it would make sense to think about a criterium to chose the default value of optimize_circuit