Benchmarking two waveform readers fairly: small dumps can flip the memory result

When comparing a multi-process waveform reader (coordinator + worker processes) against a single-process reader, measure at more than one dump size. On a 10 GB dump the multi-process reader used less peak memory (2.7 GB PSS vs 4.8 GB), but on a 145 MB dump the single process won (0.16 GB vs 0.35 GB), because each extra Node worker has a fixed ~50-100 MB base. Sum PSS over all of a reader's processes from /proc/<pid>/smaps_rollup, not RSS, so shared pages aren't double-counted. Report the small-dump loss rather than dropping it: readers notice a cherry-picked table. Also keep reader-level timings (no UI) and frame-timed UI recordings in separate tables; mixing them in one row makes the comparison meaningless.