Influencers: simplify hyper-opt/cache setup

[lefnire]

Now that top x hyper-parameters are saved to DB #77 per user, no need for the complex early-stopping logic & high n_trials. The more days pass & influencers() gets run each day, the more accurate the model becomes over time anyway. So we should just be able to depend on that, forget about beating prior scores, and keep n_trials low (either static low value like 30, or dynamic low value based on number of user's days with field_entries).

• Lower xgb.py n_trials. Either dynamic based on n_field_entries (few entries high value like 300, high entries low value like 10 - since by the time they have high entry-count, hyper-opt will have gotten pretty solid)
• Remove early-stopping logic. Just depend on the small n_trials
• Use eval_metric=mape rather than mae (docs) so trials based on different good_target can be compared; and different users can be compared.