Coordinates are delta encoded. That means the first element corresponds
to the first coordinate and all the following elements are the
respective diffs. The first element is potentially much larger than
the rest and can cause excessive integer packing due to overflows.
This change stores that first element separately.

Not sure why this doesn't cause an error here, but the code is invalid.

The chain naming strategy can be read in the documentation

- parameters for rotamers are derived from PDB structures by ourselves
- removed LOSSY and INFER_POSITIONS options. User can now just give a
  max error and OMF will figure out by itself the best strategy.
- Increased accuracy by storing angle differences in rotamer compression.
  That means: instead of blindly using the angles from the derived parameters,
  we sacrifice 4 bits to store differences to that optimal angle.
- CANNOT LOAD OMF FILES FROM PREVIOUS VERSIONS ANYMORE

SDF V2000 supports up to 999 atoms/bonds. For more we need to implement
the V3000 format. For now we throw an error.

The RCSB ModelServer SDF files contain counts line with only 33
characters. Although invalid, they can be read by RDKit (which only
requires 6 characters). This commit emmulates the behavior of RDKit.

- Required data is extracted from MMCifInfoBioUnit object and fed into a
  custom BUInfo object
- BUInfo object can be serialized into bytestring
- Chain names in biounit have following naming scheme:
  <idx>.<asu_chain_name>. idx -> n-th copy of ASU chain with n
  starting from 1.
- Properties are copied from whatever is stored in assymetric unit
- Chain properties copied: ChainType (CHAINTYPE_POLY etc.)
- Residue properties copied: OLC, sec struct, ChemClass (PEPTIDE_LINKING etc.),
  ChemType(AMINOACIDS etc.), IsProtein flag, IsLigand flag
- Atom properties copied: name, element, occupancy, bfactor, ishetatm

The intention of this breaking change is to strictly concentrate
on coordinates. It makes limited sense to provide biounit support
without also providing SEQRES etc. So this should all go to some
external annotation system.

enable contact scores and separate dockq and per-interface qs scores

- Use bond/angle parameter of PeptideBuilder
  (https://github.com/clauswilke/PeptideBuilder/)
  RMSDs get lower as compared to the previously used parameters
  that were extracted from Charmm36 IC data.
- Introduce internal coordinates to reprent backbone positions.
  Problem with peptide chains is error propagation. An error
  introduced at one backbone dihedral propagates along the whole
  subsequent chain. We thus compress short fragments (tri-peptides)
  that are independent of each other.

Test behavior when no results and of invalid "by"

This replaces the awkward FindCompound(... by=) which would return
a single arbitrary compound.

This allows preferentially picking a non-obsolete compound
in FindCompound by="smiles/inchi"