Run the test suite on free-threaded CPython (3.14t)

[bagowix]

Problem

Thread safety is a documented property of the breaker: a single threading.Lock
guards the two await-free critical sections in interlock/_engine.py:117 and is
released before the protected callable runs. Nothing in CI exercises that claim.

• The matrix in .github/workflows/ci.yml runs 3.11–3.14, all GIL-enabled.
• The suite is parallelised by process (pytest-xdist, -n auto), never by
  thread, so no test drives one breaker from several threads at once.
• Free-threaded CPython is officially supported as of 3.14, and it is exactly
  where a lock-scope assumption that happens to be protected by the GIL turns
  into a real race: the sliding-window mutation, the snapshot() copy
  (interlock/_engine.py:344), the auto-transition timer arm/cancel pair and
  the coordinator lane hand-off.

A library placed on every dependency call is the last place a latent data race
should be discovered by a user.

Proposal

Two independent changes:

1. Add 3.14t to the CI matrix. astral-sh/setup-uv resolves free-threaded
   interpreters from the t suffix, so this is a matrix entry plus whatever
   fallout it exposes. Keep it non-blocking (continue-on-error) only until it
   is green — the point is to make it required.
2. Add thread-concurrency tests with pytest-run-parallel (dev dependency,
   --parallel-threads=N), scoped to the paths where concurrent mutation is
   possible rather than the whole suite: test_breaker.py, test_registry.py,
   test_engine.py, test_coordination.py, test_observability.py.

Acceptance criteria

• CI runs the suite on a free-threaded 3.14 interpreter and the job is required.
• A marked subset runs each test from multiple threads and asserts the
  invariants that must survive contention: window counts add up, snapshot()
  never returns a torn view, HALF_OPEN probe caps are never exceeded, the
  registry hands out exactly one breaker per name.
• The core stays dependency-free; pytest-run-parallel is dev-only.
• Any race the new jobs expose is fixed or filed with a reproducer before the
  job is made required.
• Document in CONTRIBUTING.md how to run the free-threaded and threaded
  variants locally.

Non-goals

• ThreadSanitizer or other native tooling.
• Making free-threaded builds a supported distribution target (no wheels, no
  classifier changes) — this is about verifying an existing correctness claim.
• Performance comparisons between GIL and free-threaded runs; CodSpeed keeps
  measuring the default interpreter.