[core] Fix GeoJSONVTData ownership and life cycle

[pozdnyakov]

Before this change, the GeoJSONVTData instance was retained at the scheduled
lambda, which run on the worker thread represented by the GeoJSONVTData::scheduler
class member:

        std::weak_ptr<GeoJSONVTData> weak = shared_from_this();
        scheduler->scheduleAndReplyValue(
            [id, weak, this]() -> TileFeatures {
                if (auto self = weak.lock()) {
                    return impl.getTile(id.z, id.x, id.y).features;
                }
                return {};
            },
            fn);

It caused program termination in case self turned to be the last reference to this,
as the std::thread destructor was called from the thread it represented.

Now, only the GeoJSONVTData::impl class member is retained.

Fixes: #16101

[captainbarbosa]

A bit of feedback - it would be greatly helpful if these kinds of changes could be summarized with an easy to digest summary that can be used by the respective downstream platform teams. Otherwise, we end up trying to make a best guess on what changes mean, which leaves room for misinterpretation when we write our changelogs.

[zugaldia]

Agreed w/ @captainbarbosa above. If a CHANGELOG entry isn't possible at this stage, a proposed entry would be a good alternative. This helps expedite downstream releases, particularly in cases where we need to cherry pick changes for a patch releases. Thanks! cc: @chloekraw @tmpsantos

[alexshalamov]

@pozdnyakov could you please add changelog entry for this PR? thanks.