[jbolor21]

Description

Follow up to #3197 - This PR reduces repeated work and overhead when reading scenario progress by moving file access off the event loop to keep the backend responsive, while preserving progress counts, retry accounting, resume behavior, and updates to finished runs.

Changes:

• Adds a seed-to-unit lookup once per summary - reuses existing grouped attempts instead of searching all attack groups and results for every seed
• Uses binary search for result paging - locates cursor in sorted cached results and selects the requested page rather than scanning the full history for matching results
• Caches scorer metrics lookups - if file is unchanged then metrics are reused. Runs file checks & reads in worker thread so as not to block backend event loop
• Refresh summaries when metrics or scorer identity changes including for completed, failed, or cancelled runs

Tests and Documentation

All existing tests pass, added new tests for:

• Per-seed counts and avoiding repeated result scans.
• Page boundaries, shared timestamps, bookmarks between results, and bounded bookmark lookup.
• Metrics reuse across runs, file edits/deletion/recreation, and scorer changes.
• Non-blocking metrics reads, cancellation cleanup, and file permission errors.