Reduce peak memory consumption for NVCOMP Cascaded benchmark

Benchmark allow choosing "max" tunables only

Update README for GPUs

Script: Verify improved, inverse-transform-compatible HC access patterns