Country↔grid grouping core via cross_level_align

[Polichinel]

Epic: #31 · Depends on: #33

Background

The heart of the migration. The original loops per (country, time, target), gathers that country's grid cells + the country total, calls the scaling math, and scatters the result back (views_reporting/.../dataset_export.py to_reconciler / reconcile_pg_dataset). We do the same, frames-native and vectorised, on top of the slice-1 leaf reconcile_proportional (PR #30, views_postprocessing/reconciliation/proportional.py).

Work

• Take a cm frame, a pgm frame, and an injected (time, priogrid_gid) → country_id mapping.
• Use views_frames.cross_level_align to label each pgm row with its country; group by (time, country_id); for each group gather the cells + the matching country total, call reconcile_proportional, and scatter the result back into a new pgm frame.
• Pure numpy + views-frames; avoid per-row Python loops where vectorisable.

Acceptance criteria

• On the S0 fixture, per-country reconciled values reproduce the oracle within a tight, documented tolerance.
• Zeros preserved, output non-negative, per-draw country sums hold (inherited from the leaf).
• Returns a new pgm frame; input untouched.

Validation

• Unit tests on the grouping/scatter against hand-computed small cases + S0-fixture slices.

Notes

• cross_level_align requires a (time, unit)-keyed mapping. Our geography is time-invariant, so broadcast the static cell→country assignment over the frame's (time, unit) rows (or use cross_level_align_arrays at scale).