Write new Experiment or ExperimentAnalysis code module

[mdekstrand]

Right now RecListAnalysis is good but limited — only computes per-user metrics.

It would help standardization of evaluation procedures if we had a more coherent "analyze" (and maybe "run") tool for experiments. The first version, of course, would just be for analysis.

• Specify experiment axes instead of inferring them?
• Support global metrics
• Specify list lengths as analysis parameter
• Support metrics with additional data (novelty, etc.)
• Clean up metric interface design
• Support analysis (sig tests, CIs, distributions, etc.)
• Support results in DuckDB?

This ticket is really probably its own epic.

[mdekstrand]

The new RunAnalysis does this for a single run.