Honest partial-order support + explicit empty ranges

[adriendellagaspera]

Motivation

This PR started as "introduce range comparison for PartialOrd (partial_rcmp)". While
reviewing it, two issues surfaced:

1. Empty ranges were handled dishonestly by rcmp: the result was an arbitrary,
   representation-dependent Below/Above (e.g. 30.rcmp(45..35) → Below but
   30.rcmp(25..15) → Above).
2. The proposed partial_rcmp -> Option<RangeOrdering> is leaky: Below/Inside/Above
   is a total-order notion. Over a poset, a value can be incomparable with one or both
   bounds, and Option collapses "incomparable with the whole range" together with
   "comparable with one bound, incomparable with the other" — losing information and
   sometimes giving a misleading partial answer.

Both problems are solved by a single idea: represent a value's position as the pair of
its relationships to the lower and upper bounds.

Changes

• New core types
  • BoundOrdering — Within / Outside / Incomparable (the last only occurs for
    PartialOrd types).
  • RangePosition — the { lower, upper } pair, with:
    • ordering() -> Option<RangeOrdering> to collapse into a simple verdict when both
      bounds are comparable;
    • is_inside().
• Empty ranges: RangeOrdering gains an Empty variant. rcmp (for Ord) now
  detects empty ranges from their bounds and returns Empty, and never falsely returns
  Inside for an empty range. Emptiness is judged from the bounds, not the population of
  the type (so ..0u32 stays a regular range).
• Partial orders: new PartialRangeOrd trait with
  partial_rcmp -> RangePosition for every PartialOrd type, preserving per-bound
  information instead of discarding it in an Option.
• Tests: the partial suite is converted to the new type, the previously-empty test
  stubs are filled in, and the comparable-to-one-bound-only cases (the ones the old
  Option design lost) are added explicitly.
• Docs: README/lib examples updated; fixed the pre-existing name drift between the
  README (RangeComparable/range_cmp) and the code (RangeOrd/rcmp).

Notes

• Breaking change → version bumped to 0.3.0. RangeOrdering gaining Empty breaks
  exhaustive matches (intentional: it forces callers to consider emptiness).
• Asymmetry by design: rcmp (total order) detects emptiness explicitly, whereas
  partial_rcmp does not attempt it — over a poset, "is this range empty?" can be
  undecidable (the bounds themselves may be incomparable). For the partial case, Empty
  only surfaces via the (Outside, Outside) witness through ordering().

https://claude.ai/code/session_01CTKxSkrLTyauoiGSFo2dVS