Add dry run option that includes stats

[leondz]

Summary

Allow a dry run that reports stats on # calls, # unique prompts, # tokens in all those prompts.

This might be a command on the CLI, that overrides config with use of a test.Tracking generator which keeps track of how many prompts it sees and how many tokens that is, and perhaps returns some modulated version of the input as output to allow adaptive probes to get estimated to some extent.