feat(encoding): compress v2.3 mini-block offsets

[Xuanwo]

Stack 8 of 10 for Lance generic block sequence compression.

This makes Lance 2.3 variable-width mini-block offsets the first non-sparse adopter of generic block compression. The writer emits the generic representation only when its complete serialized form is strictly smaller; otherwise it keeps the legacy interleaved flat layout. Chunk-local scan, range, and take scheduling remain intact.

Lance 2.0–2.2 writers and readers keep their existing bytes and selection behavior. SparseLayout keeps its existing wire format and explicitly vetoes the generic path.

Benchmark

Representative 262,144-row UTF-8 workloads compare equivalent v2.2 and v2.3 mini-block encoding at ae0a49ba2 with the release-with-debug profile. The descendant measurement commit adds only documentation and the benchmark harness after this PR's runtime code.

Workload	Encoded bytes	Encode median	Scan median	Cold take median
Range, v2.2 → v2.3	5,283,840 → 4,218,880 (-20.2%)	1.512 → 6.824 ms	1.330 → 1.170 ms	249.2 → 202.3 µs
Delta, v2.2 → v2.3	10,436,608 → 9,355,264 (-10.4%)	2.650 → 10.014 ms	2.136 → 2.259 ms	232.4 → 246.3 µs

Stack navigation

• Position: 8 of 10
• Root: refactor(encoding): align block codecs with miniblock #8324
• Previous: feat(encoding): select compressed offset blocks #8330
• Next: docs(format): define generic mini-block offsets #8332
• Top: perf(encoding): measure compressed mini-block offsets #8333