Phase 13.14.GB: Dedicated sliding window aggregation

make_sliding_window_aggregate: pure aggregation using per-bin sufficient
statistics (sum/sum²/count). No regression, no design matrix.
  - Per-column NaN-aware counts
  - Bessel correction (ddof=1) matching fit path
  - Dense lookup array for O(1) coord→bin (Numba-friendly)
  - Row weights + kernel weights separated
  - Variance clamped to ≥0
  - Numba JIT inner loop with numpy fallback

make_sliding_window_aggregate_parallel: forkserver/spawn context,
raise on total failure.

Tests: 10 new (4 invariance, 1 performance gate)
  Key: agg ≡ fit path, window=0 ≡ groupby, parallel ≡ serial
  494 passed, 4 failed (pre-existing), 19 skipped, 0 regressions

Author: miranov25