Measure GPU starvation from nvidia-smi samples before tuning kernels

When a GPU pipeline's wall time looks slow, log nvidia-smi --query-gpu=timestamp,utilization.gpu,memory.used --format=csv,noheader -lms 500 during the run and compute the busy fraction (samples with util > 0, and >= 50%) over the whole wall time. In one run the GPU was non-idle for only ~1/3 of wall time with ~20% mean util, while /usr/bin/time showed the host process at 99% CPU (one thread). The bottleneck was host-side: copying decoded data back device-to-host (huge minor-fault counts in /usr/bin/time -v are a tell) plus single-threaded prep. Kernel tuning would not have helped. Fix direction: keep decoded data device-resident, do the prep/sort on the GPU, and stream double-buffered when input exceeds VRAM. Report busy fraction as a standard metric per run. Note: awk max over a field ending in ' MiB' can compare as strings; strip units and add 0 to force numeric.