doc(geography): document the DuckDB GEOGRAPHY boundary design

[estebanzimanyi]

Adds doc/geography-boundary.md as the canonical write-up of how MobilityDuck represents geodetic geography values across the MEOS ↔ DuckDB columnar boundary.

Why this doc exists

Two layers of context get mixed up regularly when geodetic results come out of a query:

1. MEOS has the closed-algebra property for geography. geog_in, geog_area, eIntersects(geog, geog), tgeog_length, tgeog_speed all take geodetic inputs, compute on the WGS-84 spheroid, and return properly-typed geodetic results — without leaving the MEOS C runtime.
2. DuckDB's bundled spatial extension exposes one logical type — GEOMETRY — that has no geodetic bit. The flag is at risk of being lost the moment a MEOS geography result becomes a DuckDB column value.

The doc explains the boundary-layer solution: register a GEOGRAPHY LogicalType in MobilityDuck (a BLOB alias whose payload is MEOS-WKB with the geodetic flag in the type tag), so the columnar engine carries the type information verbatim while every operation stays inside MEOS.

What the doc covers

• The problem in one paragraph + an ASCII boundary diagram.
• The GEOGRAPHY LogicalType registration sketch (~10 LoC at registration time).
• I/O surface — ST_GeogFromText, ST_AsText, ST_AsBinary, ST_GeogFromBinary — all thin shims over existing MEOS exports.
• Operation surface — length, area, eIntersects, etc. — every call delegates to a MEOS function that takes geodetic input and returns the correct type.
• The complete cast matrix (GEOMETRY / GEOGRAPHY / TGEOGPOINT / TGEOMPOINT), mirroring the MobilityDB-on-Postgres surface.
• TemporalParquet round-trip preservation via the footer JSON.
• Pitfalls a binding implementation must avoid (using ST_GeomFromText to construct a GEOGRAPHY value, reusing DuckDB Spatial Cartesian functions on a GEOGRAPHY BLOB, stripping the geodetic flag in Parquet output, etc.).
• Current state and the bounded pending work (~430 LoC, single PR) to register the LogicalType + I/O UDFs + casts + tests.

Where it's linked

• New paragraph in README.md near the parity-gaps pointer, so adopters land here when they look for geography semantics on the DuckDB side.
• The doc itself cross-links to doc/multi-duckdb-version.md and Discussion #913 — Temporal Data Lake RFC.

This is the doc the user asked for: "document all the DuckDB geography issue and solution properly so it can be widely available and findable in the documentation". The implementation (the ~430 LoC PR registering GEOGRAPHY + UDFs + casts + tests) is the natural next step.