Optimize RAD neighbor search and blocking loop in CodeEntropy pipeline

[harryswift01]

Summary

This PR introduces initial performance optimisations to the RAD neighbour search used in the entropy calculation pipeline. The changes reduce computational overhead in the RAD blocking loop by eliminating redundant operations, improving periodic boundary handling, and precomputing key quantities, while preserving identical numerical results and deterministic behaviour.

Changes

Precompute and reuse core quantities:

• Precompute center of mass (COM) coordinates per frame to avoid repeated evaluation
• Reuse precomputed distance arrays throughout sorting and RAD evaluation
• Introduce inverse squared distance values to reduce repeated arithmetic operations

Optimise RAD blocking loop:

• Remove redundant square root computations by using squared distances where possible
• Reuse existing distance values (r_ij, r_ik) instead of recomputing them
• Reduce repeated vector operations between triplets of molecules
• Add early skip condition to avoid unnecessary blocking evaluations

Improve periodic boundary handling:

• Replace np.where-based periodic boundary corrections with scalar operations inside Numba
• Introduce a helper function to apply minimum image convention consistently
• Reduce NumPy overhead within the performance-critical inner loop

Improve memory and execution efficiency:

• Replace Python list accumulation with a preallocated NumPy array in the Numba loop
• Move invariant calculations outside inner loops to avoid recomputation
• Reduce temporary allocations and improve cache locality within the blocking routine

Impact

• Improves runtime performance of RAD neighbour search, particularly for large systems
• Reduces overhead in the Numba-compiled blocking loop
• Maintains identical numerical outputs and deterministic ordering
• Establishes a more efficient baseline for further RAD-related optimisations