perf(encoding): benchmark generic offsets

[Xuanwo]

Stack 9 of the Generic Block Sequence Compression series.

Depends on #8046.

This layer adds representative Range and Delta corpora for encoder, full-scan decoder, and cold/warm random-take benchmarks. Each case compares Lance 2.2 legacy offsets with Lance 2.3 generic offsets and includes encoded bytes in the benchmark identifier.

Local release-with-debug results (10 samples, 1 second measurement) showed:

• Range bytes: 5,283,840B -> 4,218,880B (-20.2%)
• Delta bytes: 10,436,608B -> 9,355,264B (-10.4%)
• Range take median: cold 482us -> 400us; warm 491us -> 394us
• Delta scan/take remained in the same local range; Range full scan was slower because metadata-only offsets are materialized during decode

These short local measurements are representative smoke benchmarks, not a performance acceptance threshold.

Validation:

• cargo check -p lance-encoding --benches
• cargo bench -p lance-encoding --bench encoder --profile release-with-debug -- variable_offsets --sample-size 10 --measurement-time 1 --warm-up-time 1 --noplot
• cargo bench -p lance-encoding --bench decoder --profile release-with-debug -- variable_offsets_decode --sample-size 10 --measurement-time 1 --warm-up-time 1 --noplot
• cargo test -p lance-encoding (575 passed, 5 ignored; integration test passed; doctest ignored)
• cargo clippy --all --tests --benches -- -D warnings
• cargo fmt --all

Stack navigation

• Umbrella / integration reference: feat(encoding): add generic block compression #8002
• Previous: feat(encoding): adopt generic offsets in Lance 2.3 #8046