Cache customised united-atom axes topology for frame covariance

[harryswift01]

Summary

This PR improves customised united-atom covariance performance by caching static axes-topology metadata before frame execution. The frame-local covariance loop can now reuse cached atom-index relationships instead of repeatedly rediscovering the same bonded atom topology with MDAnalysis selections.

Changes

Cached united-atom axes topology:

• Added static axes-topology metadata for customised united-atom axes.
• Added BuildAxesTopologyNode to construct shared_data["axes_topology"].
• Cached only frame-independent topology data, including UA heavy atom indices, bonded heavy/light atom indices, UA atom membership, residue heavy atom indices, and residue UA masses.

Cached axes execution path:

• Added cached axes methods in AxesCalculator for computing customised UA axes from cached topology.
• Updated FrameCovarianceNode to use the cached topology path when available.
• Preserved the existing get_UA_axes(...) path as a fallback.
• Kept all frame-dependent values, such as positions, forces, centres, axes, torques, moments of inertia, and covariance matrices, computed per frame.

DAG integration and tests:

• Registered BuildAxesTopologyNode in LevelDAG after BuildBeadsNode.
• Added atomic unit tests for the axes-topology node.
• Added coverage for cached UA axes and cached bonded-axes paths.
• Updated covariance-node and LevelDAG tests to reflect the new cached topology dependency.
• Confirmed existing entropy output remains unchanged for the profiled benzene benchmark.

Impact

• Reduces repeated MDAnalysis topology selection work inside the frame covariance loop.
• Preserves existing covariance, force, torque, entropy, Dask, and scheduling behaviour.
• Provides a safe fallback to the previous axes path when cached topology is unavailable.

[jimboid]

A nifty optimisation. PR is complete with all required changes clear and matching the description in the PR summary.