feat(benchmarking): add benchmarking code into MLFlow models to better monitor accuracy

[ThomasHepworth]

At the moment, we run our benchmarking code manually to assess model changes and guard against regressions. This is imperfect because:

1. It relies on someone manually checking match rates against our ground-truth datasets.
2. It’s time-consuming to run more than one dataset when deploying changes.
3. We don’t have a complete history of model performance over time.

Now that we have MLflow running in GitHub, it would make sense to set up a new repo (or workflow) that automatically runs model evaluation whenever we publish a new GitHub release, and tracks accuracy there.

[ThomasHepworth]

Current guidance https://github.com/moj-analytical-services/mlops/tree/main
Example repo https://github.com/moj-analytical-services/mlops-drone-detection/tree/main
https://github.com/moj-analytical-services/mlflow-access