NB: Remove the t7603 directory before merging this in.

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 99.56%. Comparing base (c35ed72) to head (bdd9d63).
⚠️ Report is 2 commits behind head on main.

@@           Coverage Diff           @@
##             main    #7637   +/-   ##
=======================================
  Coverage   99.56%   99.56%           
=======================================
  Files        1088     1088           
  Lines       96048    96062   +14     
=======================================
+ Hits        95632    95646   +14     
  Misses        416      416           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

PR #7603 made multiplications with Paulis more general and correct, but it also introduced slow down for multiplication with simple Paulis.
I have added a temporary timing script and results to the directory t7603 to benchmark multiplications before and after #7603 and at this PR. #7603 caused about 2-fold slow down, after this PR the timings are either similar or better than before:

# all times in μs with stddev-s in parentheses
# x = X(q(0)), x1 = X(q(1)), x2 = X(q(2))

x**2 * x * x * x  72(3)   --> 169(5)   --> 48(2)
x**1 * x**4       40(1)   --> 88(2)    --> 42(3)
x1 * x2           21.8(2) --> 78(2)    --> 13(1)

+1 I was wondering about the perf impact of these changes. One other change in that PR that could be a regression is the new self.on(*qubits) in Gate, which creates and initializes a new operation.

That is needed, because GateOperation.__mul__ is done by Gate._mul_with_qubits which does not accept Operation. It seems all callers of _mul_with_qubits have an Operation object, so perhaps we could replace this with Gate._mul_operation which would not need to use Gate.on(). That would be for some other PR and we'd need some benchmarks first to make sure it is worthwhile to do.