Fully port BarrierBeforeFinalMeasurements to rust

[mtreinish]

Summary

This commit migrates the BarrierBeforeFinalMeasurements transpiler pass to operate fully in Rust. The full path of the transpiler pass now never leaves Rust until it has finished modifying the DAGCircuit. The one exception is when the label is not set then we still call MergeAdjacentBarriers in the python code of the pass.

This is the first step in the improvement of the performance of this pass. We can easily leverage multhreading to potentially parallelize the analysis portion of this pass that searches for the final operations and returns the set of indices. But this is blocked on #13219 which prevents us from accessing the PackedInstructions stored in the DAGCircuit in a multithreaded context.

This commit also fixes an issue related to shared references in the disjoint_utils module around barrier labels. The combine_barriers() function was incorrectly mutating the label by reference which wouldn't persist in the DAG, and this was causing failures after the barrier was originally generated in Rust with this pass now.

Details and comments

Fixes #12253

[qiskit-bot]

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

• @Qiskit/terra-core

[mtreinish]

I ran the dedicated asv microbenchmarks on this and it yielded:

Benchmarks that have improved:

| Change   | Before [1344cddb] <stash~1>   | After [f62a6d79] <port-barrier-before-final-measure>   |   Ratio | Benchmark (Parameter)                                                  |
|----------|-------------------------------|--------------------------------------------------------|---------|------------------------------------------------------------------------|
| -        | 17.1±0.1ms                    | 12.3±0.1ms                                             |    0.72 | passes.PassBenchmarks.time_barrier_before_final_measurements(14, 1024) |
| -        | 7.42±0.03ms                   | 5.12±0.1ms                                             |    0.69 | passes.PassBenchmarks.time_barrier_before_final_measurements(5, 1024)  |
| -        | 29.7±0.2ms                    | 8.93±0.2ms                                             |    0.3  | passes.PassBenchmarks.time_barrier_before_final_measurements(20, 1024) |

SOME BENCHMARKS HAVE CHANGED SIGNIFICANTLY.
PERFORMANCE INCREASED.

[coveralls]

Pull Request Test Coverage Report for Build 11022176994

Details

• 90 of 93 (96.77%) changed or added relevant lines in 6 files are covered.
• 22 unchanged lines in 2 files lost coverage.
• Overall coverage increased (+0.07%) to 88.832%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
qiskit/transpiler/passes/layout/disjoint_utils.py	12	13	92.31%
crates/accelerate/src/barrier_before_final_measurement.rs	73	75	97.33%

Files with Coverage Reduction	New Missed Lines	%
crates/qasm2/src/lex.rs	5	91.73%
crates/circuit/src/dag_circuit.rs	17	87.84%

Totals
Change from base Build 11019425622:	0.07%
Covered Lines:	73565
Relevant Lines:	82814

---

💛 - Coveralls

[ElePT]

Thanks for taking over this pass. It looks pretty straightforward, I only had one small question.

[ElePT]

What is the difference between doing this or simply omitting the #[cfg(feature = "cache_pygates")] comment and setting the instruction argument to new_barrier.unbind()?

[mtreinish]

When cache_pygates is necessary we need to clone new_barrier because it's used in the PyInstruction here and also again in the pycache for the PackedInstruction that we insert into the dag. I used the cfg macros here to conditionally compile this so we only clone if the cache_pygates feature is enabled and otherwise skip the clone() where it's not needed. This would work fine if I removed the macros and just had it always do new_barrier.clone().unbind() regardless of whether the feature was enabled, but it was just me micro-optimizing a little as the clone() and Drop for new_barrier as a Bound<PyAny> is just python ref counting and very cheap, but I figured I could just save the overhead of that here.

[ElePT]

Looks good!