Add device-side ETKDG conformer pruning

[Matthew-Neba]

Summary

• Adds conformer pruning for ETKDG device output. Fixes Device output Pruning for ETKDG #160.
• Keeps conformer coordinates on the target GPU instead of copying them to the CPU for pruning.
• Matches RDKit's ordered greedy pruning behavior, including heavy-atom and symmetry-aware RMSD comparisons.

[greptile-apps]

Greptile Summary

Device-resident ETKDG results now support ordered RMSD conformer pruning without copying coordinates to the host.

• Adds GPU kernels for symmetry-aware and heavy-atom-aware conformer comparison and coordinate compaction.
• Integrates pruning into the ETKDG device-output path and updates native linkage.
• Adds device-output tests covering greedy ordering, symmetry mappings, heavy-atom selection, compaction, and metadata renumbering.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
src/conformer/device_conformer_pruning.cpp	Builds RDKit-compatible atom maps, launches device pruning, and compacts retained coordinates and metadata.
src/conformer_rmsd.cu	Adds warp-level aligned RMSD, ordered greedy pruning, and retained-coordinate compaction kernels.
src/etkdg.cpp	Enables conformer pruning after device results are consolidated onto the target GPU.
src/CMakeLists.txt	Adds the device-pruning implementation and RMSD library to the ETKDG target.
tests/test_etkdg_device_output.cu	Expands device-output coverage for pruning behavior, symmetry, heavy atoms, interleaving, compaction, and conformer indices.

_{Reviews (2): Last reviewed commit: "Minor fixes + in place retained list" | Re-trigger Greptile}

[scal444]

Thanks for closing another feature gap! I'll be able to review this later today or tomorrow. Don't worry about the python build failures, I need to fix our CI from our v0.6 version bump. C++ build is the important one to look at.

[scal444]

Looks great! A few small things, and one performance suggestion that can definitely be left to a followup, or handled here.

[Matthew-Neba]

Awesome! Thanks for taking a look, will fix those issues and take another deep look before the weekend

[scal444]

Also our CI should be fixed now so if you rebase onto or merge in main, those unrelated failures should go away

[Matthew-Neba]

Applied the fixes. Also applied your suggestion to use a list of the retained conformers so that we don't waste time on those no ops; although it was done in place with the list containing conformers ( groupedConfIds) + a pointer (retainedCount) instead of allocating another buffer for the list.