Marginalia Search got a major glow-up, no cap

The rewrite uses deterministic block-based skip lists over B-trees, giving the index big rizz with mad speed gains.

Tweaking block sizes to ~128 KB tapped into NVMe SSDs’ fire read rates, boosting query perf big time.

Using io_uring for position reads was sus at first but now slaps, balancing throughput and latency.