Stop the sniffer panicking on binary content

Stop the sniffer panicking on binary content The SVG search truncated the lowercased head at byte 1024. That head comes through from_utf8_lossy, which replaces every invalid sequence with a three-byte replacement character, so on binary input byte 1024 is very often inside one — and slicing a &str there panics. Reachable from any upload of a file that is not text, which is most of them. Whether it fired depended on where the replacement characters happened to fall, so it was a panic that showed up on some binary files and not others. In the worker it meant a dead process with a panic message where a sniff verdict belonged, which is the failure dam-pipeline denies expect_used to avoid. Truncate to the nearest char boundary at or below 1024 instead. Found by transferring a nine-megabyte binary through the migration path. The regression test sweeps the starting offset rather than picking one blob, because which byte lands on the boundary depends on the content and a single input would pass over a fix that was still wrong by one.