'model_representation' no longer made sense now that it also applies to
cases where the ligand is too far. It was renamed to 'model_binding_site'
to be more accurate. 'binding_site' was renamed to 'target_binding_site'
to match.

This was a problem in CASP16 models L4004LG020_1F where all the model
ligands were disconnected but that was not picked up properly

In principle one can have the following alignment:

XXXXXXXXXXA--------
----------AYYYYYYYY

It has a 100% sequence identity! The previously implemented logic of gap
thresholds was also not very helpfil to filter out these cases as it
operated on fraction of gaps between first and last aligned column in
the alignment. That's 0.0 and thus perfect.

This commit simplifies this logic and simply checks for a sequence identity
threshold and a minimum number of aligned columns when grouping sequences
together. This should make grouping these cases together very unlikely.

If residue number alignments are enabled, one can assume that two
consecutive residues in terms of residue numbers are connected.
The conop Processor does not necessarily connect them if the bond
is considered unfeasible. This gives inaccurate results in
subsequent stereochemistry checks.
This change forces these connections if and only if resnum_alignments
are enabled

lrmsd defines the longer chain as receptor and uses it for superposition.
If both chains have the same length, the selection becomes random.
This commit implements the same random selection as DockQ v2.2.1.
This is no conceptual change of the score itself! It just makes the
two implementations more similar.

The latter is only relevant if mmcif_conform is True and triggered a segfault
before.

Dates can be NULL in the current database schema. We never use them
anyways (ie they are not read back from the database when compounds are
loaded) so it is safe to skip them if they are missing.

Reading the BIRD data from prd-all.cif.gz which doesn't contain any
compound would result in a compound lib filled with dummy compounds and
atoms, containing no useful data. This commit skips compounds with no
_chem_comp.id (key data item) and non-loop atoms without an
_chem_comp_atom.atom_id, with a warning.
The result of reading prd-all.cif.gz is now an empty compound lib.

This avoids ignoring issues in optimized builds and hard to read errors in
unoptimized builds.

This commit adds the second header line. While this line can be
substituted for a blank line per specification, its absence triggers a
warning in RDKit about the missing 3D dimension tag.

Use full backbone atoms (N, CA, C for peptide residues, O5', C5', C4', C3',
O3 for nucleotide residues) if number of mapped residues is below 3.
This has a direct impact on RMSD. The behaviour before was a random value
as the fallback transformation in these cases was an identity matrix.
While RMSD is still computed on CA/C3' only, we now apply the transformation
derived from all backbone atoms.

Optimization has been improved by sampling several window sizes. The results
come really close to LGA (avg diff of 0.16 GDT points when evaluation all
CASP15 TS models).

In the previous implementation, an error was thrown when number of positions
was lower than the window size parameter. Now, the window size is just set
to the number of positions in that case and the algorithm runs through.
Special cases of only one or two positions are handled separately to produce
sensible output.

This changes the CSV output file to list one model ligand per line,
instead of one reference ligand.

OST only reads the first model (_atom_site.pdbx_PDB_model_num) of an
mmCIF file. This commit clarifies it in the doc, and adds a warning to
the output if models are ignored.