Don't mix full-scan and windowed-read speedups when benchmarking waveform readers

When reporting how much faster a new waveform format reads than a vendor format, keep separate rows for distinct operations: (1) full scan of every value of selected signals (streaming, hashing packed bit planes once) and (2) interactive window reads (load a signal cohort per view window, materialize values for display). They can differ by an order of magnitude (e.g. ~20x vs ~1-2x on the same data) because baseline readers often amortize differently per window, and per-event vector materialization dominates windowed reads. Each row needs its own best-configured baseline doing identical work (same signals, same output hash), and a speedup from one row must never be quoted with the other's baseline.