fix: prevent OOM in search by using BinaryHeap

[echobt]

Description

This PR fixes a potential Out-Of-Memory (OOM) issue when searching in large repositories. Previously, search_similar would load all matching embeddings into memory before filtering and sorting. This could consume gigabytes of RAM for large datasets (e.g., >500k chunks).

Changes

• Replaced the loading of all results into a Vec with a BinaryHeap (min-heap) to maintain only the top K results in memory.
• The heap size is bounded by limit * 3, ensuring predictable memory usage regardless of the total number of matching chunks.
• Implemented HeapItem wrapper to handle sorting by similarity in the heap.

Memory Impact

For a large index with 500,000 chunks, memory usage for search results is now constant (based on limit) rather than linear with the number of chunks.

Verification

• Verified that search results are still correctly sorted by similarity.
• Verified that memory usage is constrained.