Use average gate fidelity in the commutation checker

[Cryoris]

Summary

The commutation checker currently uses a different metric to check when gates commute than other places in Qiskit, such as the RemoveIdentityEquivalent. This leads to strange behavior, where one gate is treated as identity in the commutation, but is subsequently not removed as other passes do not treat it as identity. This is what causes #13547.

This PR updates the commutation checker to use the same average gate fidelity and thus the transpiler to handle almost-identity gates more consistently.

Closes #13547.

Details and comments

• Still needs a reno.
• This has one small caveat: the definition of the average gate fidelity is global phase agnostic, but our commutation rules are not. I.e. we currently do not want to commute two gates, if doing that introduces a global phase. We could extend this in the future, but this is outside the scope of this PR.
• This PR increases the threshold for gates to commute (== rotation angles must now be smaller than before in order to commute). This means the #gates we transpile a QFT to increases, see the benchmark below. However, @ElePT ran benchpress and this didn't seem to cause a noticable regression overall.

Here's a small benchmark comparing transpile time, CX depth and count for the (1) the QFT, (2) an efficient SU2 circuit with log(num qubits) depth and (3) a Heisenberg Trotter evolution on a line.

[qiskit-bot]

One or more of the following people are relevant to this code:

• @Qiskit/terra-core