We use a shader with 64 bones by default to consume less uniform memory.
But if we meet bigger skeletons we batch and handle them afterwards.
In the future we need to sort the input models to have less changes for
bound buffers. Also we might want to skin up to 4 models per a single
dispatch.
