Interface GpuTreeBuilds

One promise per batch, in order. They never reject: a failed batch reports its meshes as refused.

GPU bytes the batches hold while they build (inputs, scratch, outputs), until their outputs are claimed (GpuTreeOutputs.claim) or freed. Read it again as batches finish: a capped build's changes batch by batch (the shared pool is counted once, by the build whose batch has it).

Meshes too big for any batch on this device (or for H-PLOC): they wait for tinybvh.

Slots whose geometry is read after build returns (a capped build's batches that wait in the queue upload theirs when they start): their arrays must not be moved into a worker before then.