Add Backman partial-orientation utilities and rank heuristics

[Copilot]

Implements primitives and rank-search shortcuts motivated by Spencer Backman's Riemann–Roch theory for graph orientations (arXiv:1401.3309).

CFOrientation — partial-orientation API

• is_partial(), oriented_edge_count(), unoriented_edge_count()
• sources() / sinks() — vertices with indeg_O = 0 / outdeg_O = 0
• is_acyclic() — iterative three-color DFS over the directed sub-orientation; unoriented edges are ignored, matching Backman's definition for partial orientations
• partial_divisor() — Backman's D_O(v) = indeg_O(v) − 1; matches divisor() on full orientations but does not require fullness

CFRank — new optimized=True shortcuts

• Negative-degree: deg(D) < 0 ⇒ rank = −1, skipping the EWD call
• Riemann–Roch lower bound: when deg(D) − g ≥ 1, start the search at k = deg(D) − g + 1 since r(D) ≥ deg(D) − g guarantees D − E is winnable for all smaller k. Gated so the pre-existing K − D swap heuristic is exercised only when this bound is not useful (no behavior change on that path).

from chipfiring import CFOrientation, rank

# Acyclicity check on a partial orientation (ignores unoriented edges)
o = CFOrientation(graph, [("v1", "v2"), ("v2", "v3")])
o.is_acyclic()           # True
o.partial_divisor()      # D_O(v) = indeg_O(v) - 1, no fullness required

# Optimized rank skips EWD for deg < 0 and shrinks the k-search via R-R
rank(divisor, optimized=True).rank

Tests

11 new tests in tests/test_cforientation.py and tests/test_cfrank.py covering each new method and both new optimizations.