fix: heap Max() and PopLast() return wrong element

[suykerbuyk]

Summary

• Max() returned data[len-1], which is not the maximum in a binary min-heap
• PopLast() called Remove(Len()-1), removing an arbitrary element instead of the worst
• This caused incorrect evictions during neighbor selection, silently degrading graph quality

Fix: scan leaf nodes (indices n/2..n-1) to find the true maximum.

Test plan

• Added TestHeap_MaxAndPopLast — pushes known values, asserts Max/PopLast return the largest
• Updated TestGraph_AddSearch — expected results now reflect correct eviction behavior
• go test ./... and go vet ./... pass

[suykerbuyk]

Hi @ammario — friendly ping on this one. This is the root cause of the near-zero recall I hit with the library: in a min-heap Max()/PopLast() returning data[len-1] means searchLayer evicts the wrong candidate, so good neighbors get dropped during search. This PR plus #21 (replenish hardcoding CosineDistance) and #22 (result set bounded by k instead of efSearch) together take recall from ~0–2/10 back to expected levels. All three are small, independent, and cleanly mergeable. Thanks for merging #19 — would appreciate a look at these when you have a moment.

[ammario]

Correct fix: in a min-heap the max is a leaf node, not data[len-1]. Verified logic and tests pass.