Besides raw mapping, this also contains alignments etc for further processing

Benchmarks show that full blown lDDT computation is faster up to
homo-tetramers. After that, caching from the decomposer starts to pay off.
This threshold is now hardcoded to select the optimal strategy. However,
the possibility for all-atom lDDT has been disabled to achieve a consistent
interface (Decomposer approach does not allow for all-atom lDDT).