Firecode

- Updated library dependency versions.

- Restyled copy of input in the log file.
- Added citations to the main program and external modules if they are used (references.py, embedder.py/_print_references).
- In runs with operators, t_start_run is set before running the first operator, and it is not overwritten by RunEmbedding if it was already set. This makes sure that the final time always includes the time spent executing the operators.

- CRESTNCI keyword: passes "--nci" to CREST, making it run in "NCI mode" that is applying a wall potential to prevent unconstrained non-covalent complexes from evaporating during the dynamics.
- If metadynamic conformational searches fail (CalledProcessError) at the default GFN2-XTB//GFN-FF level they are re-launched with pure GFN2-XTB (slower but more stable and works with inorganic ions).

- Added support for passing an energy threshold to mtd_search> CREST runs ("--ewin" in CREST, set with the KCAL keyword on TSCoDe).
- Dynamically adjusted energy pruning threshold above self.options.kcal_thresh in optimization_refining if needed, so that at least a percentage of structures is retained (default minimum set to 10%).

Discontinued OpenBabel FF support
- Updated prune_conformers_rmsd group criteria, avoiding group numbers that have (on average) less than 20 active structures.
- Disconnected the Openbabel Force Field calculator from the embedder, as the XTB implementation proved uniquely versatile and robust for more sophisticated manipulations. The interface is still present in the code (inside calculators) for reference, external utility purposes and potential future re-adoption for specific tasks. Updated the rest of the documentation accordingly.

RMSD pruning significative speedup, increased extent of parallelization throughout, keywords priority
- Small bug fixes.
- Added priority to keywords, that dictate their order of execution. Options are set first (priority 1), then priority 2 keywords that modify attributes and depend on priority 1 keywords (for now, just DIST).
- Added a status_dump call after generate_candidates (debug keyword).
- Changed spring stiffness for constrained optimizations to dynamic values (0.25 to 1 for FF, 1 to 2 for SE, in Eh/bohrs). Less scrambling, more accurate poses.
- Multiembeds now are run through self.avail_cpus max_workers instead of half.
- prune_conformers_rmsd is now completely compiled with numba, cached, parallelized and ~30 times faster. The old similarity-graph-based logic was also removed in favor of the removal of any structure at the first instance of a similar one. In the future, all pruning functions should work similarly.
- Cyclical embed now generates fewer candidates, as it discards them if they are too similar to others that share the same pivots and conformation ids (RMSD-based).
- Adjusted default threshold for RMSD similarity from 0.25 back to 0.5 A (benchmarked to retain all methylcyclohexane conformers with the new pruning algorithm).
- Reduced checkpoint dump frequency (embedder.options.checkpoint_frequency) from 20 to 50, as writing large (>20k) structures so often can slow down execution and clutter the logfile.
- Reversed order of this CHANGELOG.md file, to display the most recent updates on top.