Fix: forward search_buffer_visited_set_ in MutableVamanaIndex batch search

[yuejiaointel]

Bug

MutableVamanaIndex's batch search() overload in include/svs/index/vamana/dynamic_index.h constructed its search buffer with only 2 of MutableBuffer's 3 constructor arguments. This left enable_visited at its default (false) regardless of the caller's actual sp.search_buffer_visited_set_ setting — a dead constructor argument. The static VamanaIndex::search() (include/svs/index/vamana/index.h) already forwards this parameter correctly, so this was a static-vs-dynamic inconsistency: visited-set dedup could never be enabled via the dynamic index's batch search path, even when explicitly configured.

Fix

--- a/include/svs/index/vamana/dynamic_index.h
+++ b/include/svs/index/vamana/dynamic_index.h
@@ -534,8 +534,10 @@ class MutableVamanaIndex {
             threads::StaticPartition{queries.size()},
             [&](const auto is, uint64_t SVS_UNUSED(tid)) {
                 size_t num_neighbors = results.n_neighbors();
-                auto buffer =
-                    search_buffer_type{sp.buffer_config_, distance::comparator(distance_)};
+                auto buffer = search_buffer_type{
+                    sp.buffer_config_,
+                    distance::comparator(distance_),
+                    sp.search_buffer_visited_set_};
 
                 auto prefetch_parameters = GreedySearchPrefetchParameters{
                     sp.prefetch_lookahead_, sp.prefetch_step_};

Correctness verification

For both scenarios below, verified bit-identical top-100 result ids before vs. after the fix: 0 mismatches across 5000 queries in each case, checked via direct label-id diffs before trusting any performance number.

Performance impact (measured on dbpedia-1536-900K, deg32/cw200 Vamana graph vs. HNSW M16/efc200, AWS, matched recall)

• fp32 (gcc-11/13, 16 threads): two independent measurements — +24.73% QPS and a follow-up clean re-test with zero instrumentation confound, +27.08% QPS. Residual gap to HNSW after the fix is ~7.45% (down from ~30-40% before the fix).
• fp16 (gcc-13, fp16-stored): fix alone gives +19.6% to +21.7% QPS, confirming the bug/fix is dtype-independent. Before the fix, SVS fp16 was roughly at HNSW parity (+0.75% to +2.87%); after the fix, SVS pulls ahead of HNSW by +21.0% to +24.6% at matched recall.

Tests

• Targeted: ./tests/tests "[dynamic_index],[dynamic_vamana],[graph_index]" — 80797 assertions, 6 test cases, all passing.
• Broader run — 113126 assertions, 8 test cases, all passing.

[rfsaliev]

This PR duplicates changes made in #364.
There is also the fix which avoids reallocations in case if search config's window_size/buffer_capacity is less than K