Correlate test metrics with parameter recovery

Running parameter estimation is the only sure-fire way to guarantee that DeepClean is, at the very least, not introducing noise into the strain or removing astrophysical signals. However, achieving statistically significant estimates of its impact to PE is time consuming and would ultimately bottleneck both offline and production deployments. What we need is to run PE on multiple DeepClean versions at various levels of performance (either across HP values or across epochs in a single run), and use those metrics to compare to metrics that we can compare quickly like ASDR and find things we can measure easily that we're confident have a high correlation with PE recovery. We'll still want to run PE for new architectures or optimizers we want to take into production, but having a proxy will allow us to iterate more quickly both offline and in production.

The issue names no files, tests, or entry points. Start by locating the parameter-estimation and ASDR metric workflows, then determine how multiple DeepClean versions, hyperparameter levels, or training epochs can be compared; done means identifying a quick metric that correlates reliably with parameter recovery.