CM/PGM → PredictionFrame adapters

[Polichinel]

Epic: #31 · Depends on: #32

Background

Phase 1 added to_prediction_frame() for pgm data (views_postprocessing/unfao/frames.py, SpatialLevel.PGM, unit = priogrid_gid). Reconciliation also needs the cm side (SpatialLevel.CM, unit = country_id) and must return a reconciled pgm frame without mutating its input (de-mutation; views-reporting register C-184).

Work

• Extend the frame adapters so cm prediction data → PredictionFrame at SpatialLevel.CM (unit = country_id), reusing the existing pgm path.
• Keep the boundary thin (pandas→numpy only where the input is still a DataFrame); one concept per file.

Acceptance criteria

• cm and pgm PredictionFrames build from the S0 fixture's inputs.
• views_frames.conformance.assert_frame_contract passes for both.
• Identifiers round-trip to (month_id, country_id) and (month_id, priogrid_gid).
• Inputs are not mutated.

Validation

• Unit tests alongside tests/test_views_frames_conformance.py.

Notes

• Reuse views_postprocessing/unfao/frames.py:to_prediction_frame; the cm variant differs only in level/unit.