lavapipe: unbreak push descriptor templates

the stride/offset calculation on this was all wrong, so just unroll the template before enqueuing to present a uniform stream of descriptors without trying to copy potentially megabytes of data

also add a new zink job to provide coverage for this codepath