Associate parameters with stopping criterion to allow stateful CLI parsing and improved diagnostics

The custom stopping criterion introduced in #151 stores all parameters for all criterion implementations in a single map. This requires that all parameters must be uniquely named across implementations along with degraded error handling capabilities while parsing CLI arguments.

Each instance of stopping_criterion should hold its own params map, and the CLI parser should check that any arguments relating to these criterion are valid for the selected implementation.

The existing stopping criterion parameters are all documented as follows: Applies to the most recent --benchmark, or all benchmarks if specified before any --benchmark arguments. The new stopping criterion framework should support this pattern as well to allow individual benchmarks to be run with different tolerances.

Start by locating the stopping_criterion implementation and the CLI parser entry point. Trace how parameters are currently stored and applied to benchmarks, then define the validation path for the selected criterion. Done means each criterion owns its parameters, invalid arguments are diagnosed clearly, and per-benchmark or global tolerance scoping works as described.