[projections] Adaptive geometry resampling for alternative projections

[mourner]

Closes #10756. Implements adaptive geometry resampling when reprojecting lines and polygons into non-mercator projections, so that segments that were straight in Mercator but become curved when projected have the least amount of additional points to represent the geometry accurately.

In the demo, for a low-zoom Albers USA view, we would generate ~3 times the amount of original points. With this code, it adds under 3% instead, while not impacting anything visually.

The logic is similar to Douglas-Peucker simplification but works in reverse — if a midpoint of a curve is far enough from the midpoint between two curve endpoints, we add a point in the middle and recurse to two halves. The resulting code is relatively compact and seems pretty fast too, although we might want to eliminate recursion some time later to make sure that it never hits maximum call stack error.

Launch Checklist

• briefly describe the changes in this PR
• write tests for all new functionality
• post benchmark scores
• manually test the debug page

[ryanhamley]

Overall this looks really good. I tried integrating it with my branch and Mercator and other projections work as expected. I made a few suggestions but this is great.

[ryanhamley]

These two functions don't need to be declared in the loadGeometry method

[mourner]

It was initially a closure because it captures canonical, featureExtent and cs (tile transform), all of which would have to be moved to arguments in a quite a verbose way if we move it out. We could do it but it's also not a bottleneck.

[ryanhamley]

Flow doesn't like this because it thinks that addResampled can be called with prevMerc and prevProj as undefined variables. I added if (!startMerc || !startProj) return; at the start of addResampled to get around that but maybe there's a better way.

[ryanhamley]

In my PR, I had to delete this to get unit tests to pass. Instead of returning an empty array, we need to retain the original logic if canonical is undefined/null and avoid reprojecting unnecessarily. We can refactor reproject to do this without changing the rest of the logic in loadGeometry.

function reproject(p, featureExtent, canonical) {
    let x, y;
    if (canonical) {
        const cs = projection.tileTransform(canonical);
        const z2 = Math.pow(2, canonical.z);
        const lng = lngFromMercatorX((canonical.x + p.x / featureExtent) / z2);
        const lat = latFromMercatorY((canonical.y + p.y / featureExtent) / z2);
        const projectedLngLat = projection.project(lng, lat);
        x = (projectedLngLat.x * cs.scale - cs.x) * EXTENT;
        y = (projectedLngLat.y * cs.scale - cs.y) * EXTENT;
    } else {
        const scale = EXTENT / featureExtent;
        x = Math.round(p.x * scale);
        y = Math.round(p.y * scale);
    }

    return new Point(x, y);
}

[ryanhamley]

nit: TODO comments cause lint to fail 🤷

[mourner]

@ryanhamley thanks for the review! Addressed in the commit above. I didn't look into Flow compliance because I thought we would land your PR that addresses it first, and then rebase mine making sure it keeps everything green, but we can do either way if rebasing on your side isn't too much trouble.

[ryanhamley]

Looks great!