Performance regression: decode_compressed/zstd_strings_10cols (~18% slower)

The lance-bench run for Lance `b29e602fd781e2076e86d87b5d425bb4991abbb0` detected a high-confidence compressed decoding regression: - Benchmark: `decode_compressed/zstd_strings_10cols` - Change: +18.09% slower (last 4 results versus all older results) - p-value: 0.007921 - Run: https://github.com/lancedb/lance-bench/actions/runs/30317331277 Recent Lance commits before the tested SHA include encoding work in `e1dbd14` (`perf(encoding): cache complex-all-null levels in RLE run form and drain run-by-run`) on 2026-07-25, although that commit is not confirmed as the cause. Suggested reproduction: run this criterion benchmark at `b29e602`, compare with the result immediately preceding the recent four-result window, and bisect the intervening commits.