nd-arrow: broadcast geometry encoding classifier

[robinskil]

Part of #276 (Phase 1).

Goal

A pure function in beacon-nd-arrow that classifies, from broadcast geometry alone, how each output column of a flattened ND batch should be encoded.

Signature (sketch)

classify_encoding(variable_dims, broadcast_target_shape, flatten_order) -> Vec<ColumnEncoding>
// ColumnEncoding = Flat | RunEndEncoded | Dictionary

Rules

• Outer / slow-varying broadcast dim → RunEndEncoded (run length = product of inner dims).
• Inner / fast-varying cyclic broadcast dim → Dictionary (low cardinality, short cycles in row-major).
• Non-broadcast data column → Flat.

Notes

• Exact, geometry-derived — no statistical/cardinality detection.
• No DataFusion dependency (keep beacon-nd-arrow arrow-only).

Acceptance

• Unit tests over known geometries (single/multi broadcast dims, mixed ranks, chunked shapes).