Fix overhead in test report for large test counts

[std-max]

After every test, print_latest_status() calls _split(), which iterates over all accumulated results just to display the P=/S=/F= counters in the progress line.
This makes per-test bookkeeping O(n) and cumulative bookkeeping O(n^2), and since it runs inside _stdout_lock while holding the GIL, worker threads progressively block on it.
It is particularly visible on large runs: throughput drops and CPU utilization collapses well before the end.

This fix uses three running counters and reads them directly instead of calling _split().
The other two callers of _split() are left untouched since they need the actual result lists and only run once at the end.