Enable HNSW index arrays with transparent huge pages

[TamimEhsan]

This PR enables hugepage if the device setting is set to madvise mode. The groundwork was already there in the library.

data_level0_memory_ is ~1.28 GiB and every candidate visit is a random access into it. With THP in madvise mode a plain malloc gets no huge pages, so that array sits on ~335,000 4 KiB pages against a ~2048-entry STLB — nearly every visit misses and takes a page walk. The hugepage size is set to 2MB.

The gain exists because this host is in madvise mode. On a host set to always it is already there without the patch; on never, MADV_HUGEPAGE is ignored and it fallbacks to original behaviour.

Benchmarks

dataset: GIST1M, topk = 10

EF	Original QPS	Hugepage QPS	Recall	Gain (%)
60	7,903.70	8,608.78	0.8075	8.92%
150	3,786.11	4,198.25	0.9080	10.89%
300	2,024.44	2,127.25	0.9532	5.08%
500	1,274.77	1,327.83	0.9677	4.16%
1000	624.46	670.49	0.9787	7.37%
2000	313.69	327.43	0.9825	4.38%