Use buffered streams for hnsw index save and load

[omninguyen]

Description

Wrap the save()/load() streams in BufferedOutputStream and BufferedInputStream.

Benchmark

save()

Nodes	File size	Before	After	Speed-up
100	0.18 MB	18.9 ms	2.0 ms	9.5×
1,000	1.82 MB	180.0 ms	9.5 ms	19.0×
10,000	18.17 MB	1,469 ms	75 ms	19.6×
100,000	181.7 MB	16,595 ms	1,078 ms	15.4×
1,000,000	1.82 GB	200,931 ms	14,517 ms	13.8×

load()

Nodes	File size	Before	After	Speed-up
100	0.18 MB	15.4 ms	1.7 ms	8.9×
1,000	1.82 MB	151.6 ms	9.9 ms	15.4×
10,000	18.17 MB	1,398 ms	93 ms	15.1×
100,000	181.7 MB	15,347 ms	1,658 ms	9.3×
1,000,000	1.82 GB	168,589 ms	19,795 ms	8.5×

Each node is an indexed vector; each file is the resulting hnsw.dat; each row in benchmark uses dim = 384 with 32 neighbours per node. Environment: JDK 8, 4 vCPU, 15 GB RAM, Claude Code Cloud.

[omninguyen]

@tvd12 Hi, can you review? Thanks.

[tvd12]

@omninguyen đây là một cải tiến đáng giá, nhưng hãy để mình xem để buffer bao nhiêu thì phù hợp nhé

[tvd12]

LGTM