feat(table): emit geo bounds in manifest DataFile

[tanmayrauth]

Compute geometry/geography bounding boxes from WKB during writes and emit WKB single-point lower/upper bounds into the manifest DataFile, which is what Iceberg pruning consumes. arrow-go cannot emit Parquet column-chunk GeospatialStatistics, so bounds are computed in iceberg-go.

Also fixes a panic when writing geo columns through the stats path, and lets geo bounds round-trip via LiteralFromBytes.

Closes: #992

[zeroshade]

Thanks for adding geo bounds support. I think this needs changes before merge:

• Geometry/geography manifest lower_bounds/upper_bounds are being serialized as WKB point payloads, but Iceberg v3 geo single-value bounds use concatenated little-endian float64 coordinates (XY=16 bytes, XYZ=24, XYM/XYZM=32 with XYM Z=NaN). This will put WKB headers in manifests and break readers expecting coordinate bytes.
• Geography bounds are accumulated as raw vertex min/max without type awareness, edge interpolation, or antimeridian wraparound handling. For geography, the correct longitude interval may wrap with xmin > xmax, so these bounds can be incorrect unless computed geodesically/type-aware or omitted when unsafe.

[tanmayrauth]

@zeroshade Thanks for both good catches. Switched manifest bounds to the spec coordinate encoding and made geography omit bounds for now.

On table/internal/geo_codec.go (the encodeGeoBound/Bounds area):
Fixed — bounds are now the Iceberg geospatial single-value serialization (little-endian float64 X, Y[, Z][, M]: 16/24/32 bytes, XYM writes Z=NaN), not WKB. LiteralFromBytes for geometry/geography now validates the 16/24/32 length instead of treating the payload as opaque WKB, and the tests assert exact byte lengths and the decoded coordinates (including the XYM NaN slot).

On table/internal/parquet_files.go (the AddWKB accumulation loop):
Made accumulation type-aware: the accumulator now knows geometry vs geography (from the Iceberg field type), and geography bounds are omitted entirely. Since geography edges are geodesics, vertex min/max isn't a safe box — latitude can bulge past the endpoints and longitude may need to wrap the antimeridian — so rather than emit possibly-wrong bounds that silently prune valid rows, I leave geography unbounded until we add geodesic/antimeridian-aware computation. Geometry (planar edges) still gets exact bounds.