Fix jsonl read bottleneck

[JojiiOfficial]

Load dataset's queries at startup into memory, so they don't bottleneck during search.
I ran into this during my experiments despite using a fast nvme.

The Issue

Jsonl reader costs a full File::open plus a full serde_json parse for each query. On H&M dataset for example, queries are quite large ~45kb.

Benches

bfb with -p 12 -t 12

	client time per request	concurrent achieved (of 12, median)	searches/s
dev	1.005 ms	7.0	5,478
PR	0.291 ms	10.7	15,092

Difference in real world benchmark (two runs, same server and search setup).

binary	median qps	avg server time	concurrency achieved*	recall@10
dev	4,757	0.63 ms	~3.0 of 32	0.92079
PR	7,250	4.12 ms	~29.9 of 32	0.92079

* qps x latency

Costs

Loading it at startup is expensive and can take a few seconds depending on disk.
For reference, H&M query set takes 850ms on my machine, but my storage is on the faster side.
I still think it's better to unconditionally (= not based on any configurations) fix this bottleneck so we can rely on measurements being correct.