feat(geo-layers): TerrainLayer GlobeView + grid tesselator

[charlieforward9]

Summary

Makes TerrainLayer render correctly on GlobeView and adds a grid tesselator that stays valid across projection toggles.

Three related changes:

1. tesselator: 'grid' (opt-in) — emits a fixed-resolution lng/lat/elev mesh in-process (no worker). The default 'auto' path keeps using Martini/Delatin via @loaders.gl/terrain. Grid rows are uniform in Mercator-y so they line up with heightmap rows at any latitude (no polar sampling warp). Because vertices are lng/lat, the same cached mesh renders correctly on both MapView and GlobeView — no re-tesselation on projection toggle.
2. Coordinate system selection in renderSubLayers — picks LNGLAT for the grid tesselator (vertices are lng/lat) and for the legacy path on GlobeView (where GlobeViewport.projectFlat is identity, so bounds land in lng/lat degrees). Uses CARTESIAN for the legacy path on MapView. Previously, legacy cartesian meshes landed inside the sphere on GlobeView and were invisible.
3. Horizon culling in tile-2d-traversal — the frustum test accepts tiles on the far hemisphere of the globe; without a horizon check, low-zoom traversal fans out across the whole globe and loads many tiles that can't be visible. A point on the sphere is visible from camera C iff dot(P, C) > |P|^2 — applied to the tile-vs-camera closest point (with lng wrap so the clamp is angular, not Euclidean).

Two quality touches on the grid path:

• Elevation samples clamped to [-500, 9000]m. Terrain-RGB encodes into 24 bits across three channels, so a single corrupt pixel (PNG decode artifact, overzoom noise, partial load) decodes to millions of meters and renders as a specular star. The clamp covers the full range of real terrain.
• Edge-vertex skirts clamped to at most 1% of the in-tile grid step so the seam slope stays under ~0.6 degrees at any zoom. The default meshMaxError * 2 (~8m) over a ~0.6m grid step at z=21 produced a near-vertical cliff at every tile edge.

Default material changes from true (PBR with specular) to a matte configuration. Specular on terrain meshes catches as star-shaped glints on skirt edges and elevation discontinuities; matte reads as a fluid landscape.

Why

Grid tesselation makes terrain cheap enough to keep cached on projection toggle, and solves the polar sampling warp that lat-uniform tesselators exhibit near the poles. The LNGLAT coord-system fix and horizon culling are prerequisites for using TerrainLayer under GlobeView at all.

Test plan

• MapView with tesselator: 'auto' — unchanged behavior
• MapView with tesselator: 'grid' — renders equivalently, single cache across zoom levels
• GlobeView with tesselator: 'grid' — renders correctly from z=0 through z=21+
• GlobeView with tesselator: 'auto' — renders (via the new LNGLAT branch)
• Projection toggle without rebuilding the tile cache
• No specular star artifacts on z=21+ close-ups (flat beach terrain)
• No visible tile-edge cliff at z=21+
• Would welcome a render-test comparison vs. current master on the terrain-extension example

Companion PRs

• Independent to merge, but camera interactions on GlobeView benefit from feat(core): GlobeView pitch, bearing, and cursor-anchored zoom #10243
• Unlocks feat(extensions): TerrainExtension GlobeView support #10245 which uses this grid tesselator in its example

[charlieforward9]

Superseded by #10250 (same branch re-opened under cr/ prefix, pushed directly to visgl/deck.gl).