Now we always have a warning on the deprecated doc with a link to the
new implementation; and a note on the new implementation which starts
with "this is the new implementation" rather than "XXX is deprecated".

The breadcrumbs now correctly indicate the module hierarchy and can be
used to determine imports. All the documentation for ost.* is now rooted
on the base `ost` module for more clarity.

ost.mol.alg submodules are now documented in separate pages. This should
make it clearer what to import.

This way we avoid making networkx a hard dependency on mol.alg.

parasail produces slightly different alignments for one test case

Use np.isnan instead of scipy.isnan. Starting from scipy 1.12
(or earlier), scipy.isnan was definitely deprecated. This made the
SpearmanCorrel function silently return None which triggered a failing
unit test that was reported by Andrius Merkys.

QS-score which is used as default target function for chain mappings.
However, thats protein specific. As soon as nucleotides are involved,
lDDT with an increased inclusion radius of 30A is now used
in ChainMapper.GetMapping.

It was a bit embarassing that lDDT mappings were about an order of
magnitude slower than QS-score mappings. A specialized backbone only
lDDT has therefore been introduced that uses matrix operations.
This is not a replacement of the lDDTScorer. But it doesnt need to
produce per-residue scores and doesn't need to deal with symmetries
etc.

Must be enabled at compile time. It bends around the respective
naive implementations in OST to their parasail counterpart.
On one hand its faster and on the other hand it should fix potential
suboptimal alignments which are reported in SCHWED-6266

The reason for removing MATCH and IDENTITY is to only have matrices
which are also available in the parasail library. This simplifies the
usage of parasail as a drop-in replacement. To be honest: I doubt that
anyone used these matrices.

The trace directionality for the internal coordinate system was
established in a recursive fashion. This likely caused filling up
stack memory for very large polymers with large recursive depth.
The fix implements the same functionality in a non-recursive
fashion.

They have the exact same quaternary structure in the end... This definition
has been generalized to: if there are no contacts observed, QS-score is 1.0
if the scored structures have the same stoichiometry, 0.0 otherwise.
This becomes a bit more abstract for higher order complexes but in principle
they still have the exact same quaternary structure if they match in
stoichiometry but have no single contact.

The score is simply not valid and returning 0.0 as done before is incorrect.
Delegates further treatment of the problem to the caller which might want to
say: if both structures are monomer, I still give it a value of 1. Or, to
extend that definition, if I reproduced the correct stoichiometry, I still
give it a value of 1.

Parasail: https://github.com/jeffdaily/parasail

It's faster than the pairwise sequence algorithms we have in
OpenStructure. However, the main reason to check out another solution
is the fact that OpenStructure may return suboptimal alignment
results. This is a result of how the dynamic programming table is built.
If for a certain cell in the dynamic programming table a match state and
insertion state score equally, match is preferred. There is a chance
that we would get a better score later on if we would stick with the
insertion state (gap open penalties vs. gap extension penalties).
Not sure if this will be fixed in the near future.

This is in order to be consistent with the behavior of the Scorer

This avoids clashes between lDDTScorer classes in ost.mol.alg
and ost.mol.alg.lddt.