gaussianParseConcurrency

Concurrent splat-parse cap. Each in-flight SPZ parse holds tens of MB transient (compressed buffer + decoded position/scale/rot/SH FloatArrays before GPU upload), so the general fetchConcurrency would peak at ~2 GB on a splat tileset. Mesh fetches on the same layer keep fetchConcurrency's budget — only the Gaussian branch gates here. Setter rebinds the semaphore.