Improve Coconut top-k seeding and leaf reads

[sophisid]

Coconut exact top-k previously started with an empty heap, ran queries serially, and reopened leaf files for each candidate vector. This change seeds refinement from distinct records near the query's sortable-SAX key, honors the requested query workers, and reads complete candidate leaves into private query buffers. Exact refinement still considers every unseeded SAX summary and the live insertion buffer.

Top-k and range now share checked leaf reads, including partial final leaves; missing or truncated reads raise an error instead of producing incomplete results. Contiguous build inputs are borrowed only while writing the index. Regression tests cover FAISS ground truth, streaming inserts, worker counts, ties, directory endpoints, oversized k, input lifetime, invalid configuration, and failed reads.

The README and pinned upstream audit explain the remaining storage differences: the key directory is resident, and inserted records remain in an unmerged buffer. This does not implement Berkeley DB storage or Coconut-LSM flush/merge. /scripts/experiments/ is added to .gitignore; experiment scripts, datasets, and unrelated MESSI changes are excluded.

Validation:

• Release build of the C++ library and Python bindings with BUILD_COCONUT=ON.
• CTest: test_coconut_topk_py and test_streaming_py passed (five Coconut checks and four existing streaming checks).
• Release library build with BUILD_COCONUT=OFF, with MPI, CUDA, FFTW-backed algorithms, and C++ FAISS disabled.
• git diff --check passed.