v0.3.1

A performance release with no API changes. Results are unchanged, except
that symmetry-reduced factorization caches may return different (equally
valid) witnesses.

Changed

• Faster sequence comparisons, which no longer expand the term lists, and
  faster catalogue divisor queries.
• Atom catalogue enumeration extends only zero-sum free prefixes and needs no
  separate atom tests; catalogues are tens to thousands of times faster, for
  instance under a second instead of about half a minute for C₃³.
  is_atom() checks a single set of subsequence sums instead of enumerating
  subsequences.
• Automorphism groups of image-table automorphisms, such as those built by
  AutomorphismGroup.closure(), compute images, orbits, orbit minima,
  stabilizers and transporters with integer keys instead of building image
  sequences. Small orbits are traversed through the generators, larger ones
  found by scanning the group. Canonical forms of generic sequences are 25
  to 70 times faster, stabilizers and orbit classification 3 to 9 times.
  FactorizationCache with symmetry benefits accordingly; for such groups it
  transports witnesses with the smallest automorphism realizing the
  canonical form, so returned witnesses may differ from 0.3.0.
• The factorization solver packs remainder multiplicities into single integers
  with guard bits, so testing whether an atom fits and computing the remainder
  take one subtraction. Together with the faster atom tests, length sets are
  about four to five times faster.
• Zero-sum constant searches with symmetry reduction find canonical forms
  without scanning the whole group at every step; η(C₃³) takes under a second
  instead of about four.
• sum() and subsequences() no longer coerce terms that are already
  canonical.

Added

• benchmarks/benchmark_operations.py times catalogue enumeration, atom
  tests, length sets, automorphism operations, and zero-sum constants over
  several kinds of groups.

Full changelog: https://zero-sum-sequences.readthedocs.io/en/latest/changelog.html