Segment benchmarks by core/support #2980

Follow-up to the position dimension left open in #2969 (point 2), now that position_est (#2974) and the brackets (#2972) are both in.

Problem: benchmarks are written from api data in counts.ts, and the position estimator needs parsed data (gold_t, lane_role), so it can't run there.

So I measured what the api data alone can do, on the same 100 pro matches from the #1590 study: ranking each team by final gold_per_min gets the exact 1-5 position right only 58% of the time, but it separates cores from supports on 99.4% of players (924/930).

So the proposal: segment the bracket benchmarks by core/support instead of 1-5. That's the split that actually changes the numbers per hero (a support Mirana and a core Mirana shouldn't share a distribution), it works at full volume with no parse dependency, and it only doubles the bracket keys instead of x5. Exact 1-5 could come later from parsed matches only, if that sample turns out dense enough.

Happy to put up the PR if this sounds reasonable.