Multi-member ReCom

[peterrrock2]

Summary

Don't worry about the docs failing for now.

Restores ReCom support for plans whose districts elect different fixed numbers of members, rebuilt
on the current GerryChain contracts. Adds a parallel proposal API, a per-member population
constraint, documentation, and reproducible asset generation, leaving the single-member API
untouched. Stacks on tech-debt/finish-fixing-types.

Why

Multi-member ReCom existed in a separate branch before the RustworkX migration but needed migration
following the rewrite.

Each district label carries a positive integer member count, and its population target is
pop_target * members_per_district[label], where pop_target is the target for a single member.

Changes

• Added multi_member_recom, build_multi_member_recom_proposal_fn, and the non-instantiable
  MultiMemberReCom namespace (four non-reversible variants) in
  gerrychain/proposals/multi_member_tree_proposals.py, exported from gerrychain.proposals.
• Added epsilon_tree_bipartition_multi_member, a proposal-level wrapper that computes the feasible
  population interval for the two merged labels, converts it to the scalar symmetric form
  bipartition_tree already accepts, and calls it with single_district_cut=True. An empty
  interval raises PopulationBalanceError immediately rather than retrying trees or reselecting
  pairs, since a different tree cannot make the numeric interval feasible.
• Added within_percent_of_ideal_population_per_member, which bounds each district's population per
  member around the plan-wide ideal (total population / total members).
• Added a worked multi-member section to docs/user/recom.ipynb using the per-member constraint.
• Added a make docs-recom-assets target and extended docs/generate_recom_assets.py to regenerate
  every static Gerrymandria plan image reproducibly; the regenerated assets are included.

Testing

• Tests pass locally
• Added/updated tests
• Type checks pass locally

tests/test_proposals.py covers exact and epsilon-tolerant splits, per-label bounds preserved
across a seeded chain, mixed-type labels, seeded-trajectory reproducibility, all four namespace
variants matching direct calls, region-surcharge plumbing, pair reselection, infeasible-interval
PopulationBalanceError, and the full validation matrix. tests/constraints/test_validity.py and
tests/typing_assertions.py cover the constraint and its typing contract. Verified with
make test, make lint, make type-check, and make docs.

Reviewer Notes

• Member counts are fixed to district labels for the whole chain; the mapping keys must equal the
  partition's labels exactly, and every count must be a positive integer (booleans rejected).
• The design deliberately avoids touching _PopulatedGraph, the cut finders, or the tree protocols.
  Unequal-target support is confined to the new proposal-level wrapper.
• The multi-member caveats are documented beside the example: the equal-district population
  constraint is not appropriate, pair selection is not seat-weighted, and there is no reversible
  variant (reversible multi-member ReCom needs a separate mathematical design).
• Random multi-member seed generation is left for a future release: users can just merge
  single-member districts for now.