test(core): add random-access skiplist benchmark

test(core): add random-access skiplist benchmark

Addresses review feedback from @nymius: existing benches use fixed
targets, which can land favorably or unfavorably relative to skip
pointer positions and don't reflect real query patterns. The new
bench draws 256 targets from a deterministic xorshift sequence and
runs both a skiplist-enhanced get() and a plain linear walk over
a 100k-node chain, so the same query stream exercises both paths.

Also changed benchmarks to reflect checkpoint full chains and removed
unhelpful benchmarks.