feat(zvec): add DiskANN benchmark support

[gaowayne]

Summary

Add DiskANN support to the Zvec backend, reusing VectorDBBench's existing dataset, loading, search, recall, latency, and result-reporting workflows.

HNSW remains the default.

Changes

• Add ZvecDiskANNIndexConfig and support restoring DiskANN configurations from saved results.
• Introduce --index-type diskann with separate construction, search, and PQ parameters.
• Map configurations to native DiskAnnIndexParam and DiskAnnQueryParam.
• Reject algorithm-incompatible CLI options.
• Protect existing non-DiskANN collections from DiskANN rebuild operations.
• Validate persisted schemas, per-document insert results, document counts, and index completeness.
• Support search-only runs against existing DiskANN collections without rebuilding.
• Avoid calling close() after successful collection destruction.
• Add regression tests, an opt-in Linux native smoke test, and usage documentation.

Validation

Validated on Linux with Python 3.12.3:

• Adapter/configuration/CLI tests: 35 passed, 1 skipped.
• Opt-in native DiskANN test: 1 passed.
• Native test covers insertion, optimization, close/reopen, search, and rebuilding.
• Cohere 10M build completed with 10,000,000 documents and 1.0 index completeness.

Cohere 10M Search Results

Configuration: 768 dimensions, COSINE, R=64, build list=100, PQ=96,
search list=300, K=100, nq=1, and IDs-only responses.

Metric	Result
Insert + optimize duration	1,657.05 s
Highest tested QPS	1,650.32 at 20 concurrent processes
Recall@100	0.8941
NDCG	0.9153
Serial P99 latency	10.5 ms
Concurrent P99 latency at 20 processes	15.922 ms

Concurrent search was tested at 12, 14, 16, 18, and 20 processes,
with a 300-second measurement window per configuration.

Scope and Limitations

• Uses Zvec's native C++ DiskANN implementation, not Microsoft Rust DiskANN.
• No changes to the shared benchmark runner or native indexing algorithms.
• The existing task assembler determines the effective metric from the dataset.
• Requires a DiskANN-enabled Zvec SDK with Collection.close() support.
• These results represent one configuration, not a controlled equal-recall comparison with HNSW.
• Memory consumption and storage I/O benefits have not yet been measured.

[gaowayne]

/approve

[sre-ci-robot]

[APPROVALNOTIFIER] This PR is NOT APPROVED

This pull-request has been approved by: gaowayne
To complete the pull request process, please assign xuanyang-cn after the PR has been reviewed.
You can assign the PR to them by writing /assign @xuanyang-cn in a comment when ready.

The full list of commands accepted by this bot can be found here.

Details

Needs approval from an approver in each of these files:

• OWNERS

Approvers can indicate their approval by writing /approve in a comment
Approvers can cancel approval by writing /approve cancel in a comment

[gaowayne]

/assign xuanyang-cn