feat(frontend): WebGPU point layer for large GeoParquet datasets

## Problem

Large point GeoParquet datasets (>1M points) choke MVT rendering via tipg — tiles get huge, render stalls, and interactivity suffers. Today CNG always serves vectors as MVT regardless of geometry type or count.

## Proposal

Add an optional WebGPU-backed scatter layer for point-heavy GeoParquet datasets, inspired by [Geospatial Atlas](https://github.com/do-me/geospatial-atlas) (fork of Apple's Embedding Atlas). Their engine handles ~200M points in the browser.

When a dataset is detected as points-only and exceeds a threshold (e.g., 100k features), offer WebGPU scatter as the render mode instead of MVT.

## Scope

- Detect point-only GeoParquet at ingestion; persist a `render_hint` on the dataset row
- Add a new frontend layer type that reads GeoParquet directly via DuckDB WASM (or a direct HTTPRangeReader)
- Fallback to MVT for browsers without WebGPU (WebGL 2)
- Surface as a toggle in the map view

## Out of scope

- Replacing MVT for non-point or small datasets
- Backend changes to tipg

## References

- https://github.com/do-me/geospatial-atlas
- https://apple.github.io/embedding-atlas

Start by tracing the dataset ingestion path and map view, then review the Geospatial Atlas references and compare DuckDB WASM with a direct HTTPRangeReader. Done means point-only, large GeoParquet datasets persist a render hint, expose a WebGPU scatter toggle, fall back to MVT with WebGL 2, and leave small or non-point datasets unchanged.

Written by the indexing model from the issue text.