Better heuristic for when to optimize target() metrics

[Zac-HD]

Currently, we run all of our optimizer routines exactly once, starting at (or just after) half of our max_examples budget. This works pretty well for the default max_examples=100, but I've noticed increasing use of Hypothesis with very large example budgets (and target()) for one-off searches for some example - and we can do much better for this workflow. Specifically, we should:

• better support interleaving of ordinary generation with targeted optimization; e.g. "run an optimize step, then generate that many examples, then the next optimize step...". This still aims to spend up-to-half our total examples optimizing.
• pick a smoother heuristic for when-to-optimize. Initial proposal: maintain current behaviour for max_examples < 1000; for max_examples >= 1000 use the incremental mixture instead, starting after 200 examples generated by our usual means.
• for long runs, we might finish running the optimizer. How should we decide to start a second full pass? We don't want unproductive re-runs (as may be implicated in Pareto-optimizer sometimes causes tests to run much more slowly #2985), but an additional run must sometimes make sense...