Cutting VCD prefixes by size: don't infer the last timestamp from a small tail read

When producing size tiers (e.g. 10/25/50 GB) as byte prefixes of one long VCD cut at timestamp boundaries, a script that finds the final cycle by reading only the last few KB of the file breaks on wide designs: at ~20 KB of value changes per cycle, a 4 KB tail can contain no '#<time>' line at all, so the cycle count comes out as 0 and the tiers are cut as header-only files.

Fixes: - Scan backwards in growing chunks until a '#' timestamp line is found (or read a generous tail such as 256 MB), and assert the count is nonzero before cutting. - Cut each tier at the byte offset of a '#<time>' line so every prefix is a valid VCD ending on a time boundary; record the offsets and cycles. - Test the cutter on a fixture whose per-cycle output exceeds the tail size; small toy VCDs won't catch this. - Store the full VCD compressed plus the offsets, so any tier can be rebuilt exactly with 'zstd -dc file.zst | head -c BYTES'.