Columnar waveform blocks: rows-per-block trades file size against short-window read latency

In a per-signal block layout (each signal's value changes chunked into compressed blocks of N rows), growing N from 256 to 16K rows shrank files to ~0.47x and made full-signal scans ~3x faster, but made short time-window reads ~8.7x slower (1K: 0.64x size, 2.2x slower windows). Point reads were unchanged. The cause: a window read must decode the whole block containing the wanted rows, so latency scales with rows per block. Keep a small default when interactive viewers read short windows; offer large blocks as an archive/scan mode. Measure warm and cold reads before deciding: cold is where block size bites hardest.