Add free-threaded Python to tox and CI, and add tests sufficient to claim compat

[bskinn]

Objective is to claim compatibility with free-threaded builds, but not (yet..?) promise thread-safety of the API and classes.

Edited ChatGPT suggestions for tests:

1. A “GIL stays off” smoke test

   Add a test that runs in a free-threaded interpreter and checks the runtime GIL state before/after import:

   assert sys._is_gil_enabled() is False early in the process
   import sphobjinv
   assert sys._is_gil_enabled() is False after import

   This catches the “a dependency re-enabled the GIL” class quickly.
   Python Free-Threading Guide

2. Multithreaded stress tests around your public API

   Even if sphobjinv itself doesn’t create threads, your users might. Add a small set of tests that:

   • start (say) 8–32 threads

   • repeatedly do parse → inspect → serialize cycles on inventories

   • use separate temp dirs/files per thread (so you’re not testing filesystem collisions unless you want to)

   What you’re looking for is: crashes, hangs, corrupted outputs, nondeterministic exceptions.

   Why: global caches / module globals / shared singletons are the number one problem in pure-Python packages.

3. Run the same multithread tests on the regular build with an ultra-low switch interval

   The guide suggests using a very short thread switch interval to flush out thread bugs even on the GIL build.

   So: in that threaded test module, set something like sys.setswitchinterval(1e-6) (and restore after). This often makes racy code fail faster.

4. A “shared object misuse” test (optional but valuable)

   Decide what you want to support, then test it explicitly:

   • If you don’t want to promise that a single Inventory instance is safe to mutate from multiple threads, document that and only test “separate instances per thread.”

   • If you do want to support it, add a test that shares one instance across threads while performing read-only operations (and possibly controlled mutations), and assert consistency.

5. Property-based / fuzz testing for determinism under concurrency (nice payoff)

   For a tool like sphobjinv, it’s natural to test invariants such as:

   • decode(encode(inv)) round-trips

   • string/JSON outputs parse back to the same semantic inventory

   Run those invariants under a thread pool to catch “rare” races.

[bskinn]

#333 added free-threaded envs to the tox matrix.