Fix positions_hours_before undercounted for open gaps at window overlaps

[tomaslink]

Summary

• Boundaries.last_boundary() picked whichever boundary starts latest, regardless of whether it had the fullest history for the vessel's actual last message. When that message falls in the ~12h overlap between two sliding windows, the later-starting window's own lookback is shorter, so the open gap created in Step three (eval_last) got a truncated positions_hours_before/positions_hours_before_sat.
• Added Boundaries.boundary_for_last_message(), which picks the earliest-starting boundary whose own last message matches the overall last message across all boundaries — i.e. the boundary with the fullest lookback, not just whichever window starts latest.
• Confirmed against real production data (gfw-int-ais-datalake.gaps_v1.raw_gaps_versioned): this recurs at every internal sliding-window boundary inside both the daily-incremental DAG and the yearly backfill runs (window period caps at 180 days per MAX_WINDOW_PERIOD_D), with undercounting sharply concentrated in the ~12h window right after each boundary and clean everywhere else.
• Ref: PIPELINE-4631

Test plan

• New parametrized case one_ssvid_with_open_gap_at_window_overlap_boundary added to tests/conftest.py's TestCases.GAP_BETWEEN_DAYS, exercised by the existing test_detect_gaps.py suite
• Full test suite passes (114 passed)
• Verified against real BigQuery data: positions_hours_before_sat corrected from 3 to 16 for the reproduced case