Unify Butina outputs for device and RDKit compatibility mode.

[scal444]

Addresses #269 , though with another API modification. There's no good way to unify fused/nonfused Butina + our previous history of return types without a clean approach.

[greptile-apps]

Greptile Summary

This PR unifies the output contracts of the matrix-based and fused Butina clustering APIs.

• Adds explicit device-resident and RDKit-compatible output modes.
• Introduces a structured device result containing cluster IDs, centroids, and cluster sizes.
• Optimizes RDKit cluster reconstruction while preserving deterministic member ordering.
• Updates tests, API documentation, usage guidance, and benchmarks for the new contract.

Confidence Score: 5/5

The PR appears safe to merge, with no outstanding correctness, security, or repository-rule issues identified.

The recent cluster-size and RDKit-conversion changes preserve CUDA stream ordering, tensor ownership, contiguous cluster mappings, and deterministic member order; no blocking failure remains.

Important Files Changed

Filename	Overview
nvmolkit/clustering.py	Introduces unified output modes and structured device results, with stream-safe cluster-size calculation and RDKit-compatible host conversion.
nvmolkit/tests/test_clustering.py	Updates existing callers and adds coverage for output modes, cluster sizes, exact RDKit compatibility, and member ordering.
skills/nvmolkit-usage/SKILL.md	Updates the Butina workflow example and explains how to consume device and RDKit output representations.
benchmarks/butina_clustering_bench.py	Selects device output explicitly so the benchmark continues measuring the intended GPU-oriented path.
docs/api/nvmolkit.rst	Adds the new output enum and device-result class to generated API documentation.

_{Reviews (2): Last reviewed commit: "Address Butina review feedback" | Re-trigger Greptile}

[scal444]

/nvskills-ci

[ohadmo]

/nvskills-ci

[ohadmo]

/nvskills-ci

[evasnow1992]

Thank you for consolidating butina outputs and addressing the comments. Changes look good to me.