ObjectPool: synchronize mutating methods with a Mutex on threaded targets

[AxGord]

Summary

lime.utils.ObjectPool mutates shared state (__pool, __inactiveObject0/1,
__inactiveObjectList, activeObjects, inactiveObjects) from get /
release / add / remove / clear / size-setter without
synchronization. When two threads drive the same pool concurrently, the
counters drift and the same instance can be handed out twice.

This PR adds a sys.thread.Mutex guarded behind #if target.threaded so
single-threaded targets (js, flash) keep the existing zero-overhead path,
and adds unit tests covering both the regular API surface and the
multi-thread regression.

The bug

Both get() and release() decrement-and-increment activeObjects /
inactiveObjects and reassign __inactiveObject0/1 without locking.
Classic interleaving:

• Thread A reads inactiveObjects > 0 → true, enters __getInactive().
• Thread B reads inactiveObjects > 0 → still true (A hasn't decremented yet),
  also enters __getInactive().
• Both threads pull __inactiveObject0, the same instance is returned to
  both callers; or inactiveObjects is double-decremented and ends up negative.

Fix

• Add sys.thread.Mutex __mutex field under #if target.threaded.
• Wrap the mutating public methods (add, clear, get, release,
  remove, set_size) with two inline helpers __lock() / __unlock()
  that compile to no-ops on non-threaded targets.
• Internal helpers __addInactive / __getInactive / __removeInactive
  are only called from already-locked public paths — they don't lock again.

One small behaviour change

In #if debug mode release() now also calls __unlock(); return; after
each Log.error so a release with an invalid object can't continue into
activeObjects-- and __pool.remove(object). Without the early return,
the function continued and corrupted pool state further past the detection.
Only observable in debug builds.

Tests

New tests/unit/src/utils/ObjectPoolTest.hx (registered in TestMain.hx):

• Eight single-threaded functional tests: get/release reuse, size cap,
  clean callback, clear, remove, size-setter pre-fill.
• One multi-thread regression testConcurrentGetReleaseDoesNotDriftCounters
  guarded by #if target.threaded: 8 worker threads × 1000 get/release
  iterations on a 4-slot pool, with a 5-second deadline in the waiter
  loop so the test fails cleanly if pool state corrupts (no CI hang).

Result on neko (un-mutexed run is from develop, run via temporary checkout):

Variant	Counters after run	testConcurrentGetReleaseDoesNotDriftCounters
unpatched	active=5, inactive=-1 (typical)	FAIL (counter drift)
patched	active=0, inactive=4 (every run)	PASS

Total: 9 tests, 23 assertations, ~11 ms on neko in both builds.

Risk

• Non-threaded targets (js, flash) — zero runtime cost. __lock() /
  __unlock() are inline and their bodies are #if target.threaded-gated,
  so every call site (get, release, add, clear, remove, set_size)
  compiles away completely. Verified by inspecting generated JS: no
  __lock / __unlock / __mutex references at any call site. Two
  empty function definitions remain in the output (~40 bytes of dead code,
  never invoked) — happy to #if-gate those away too if reviewers prefer
  zero bytes.

• Threaded targets — single-thread use pays an uncontended-lock cost.
  Benchmarked on neko (5M get/release iterations, single thread, 64-slot
  pool): unpatched ~330 ns per get/release pair, patched ~565 ns — about
  +70% on neko, where Mutex is heavy. On cpp/hl with native
  uncontended mutex (std::mutex / pthread futex) the cost is typically
  in the low-tens-of-ns per acquire/release, so the percent overhead is
  much smaller. In real OpenFL hot paths (Event pool, Graphics
  shaderBuffer pool, etc.) the call rate is ~tens-to-hundreds per frame,
  so the absolute cost stays in the microsecond range per frame even on
  neko.

• Multi-thread use is the entire point of the PR. Without the mutex,
  the test above demonstrates the pool ends up with inactiveObjects=-1
  and hands out duplicates; the patched version is stable.

• If reviewers want to keep single-thread users opt-out, I'm happy to
  add a -D lime_objectpool_no_mutex compile flag in a follow-up.

[joshtynjala]

I wonder if it would make sense to add a separate ThreadSafeObjectPool.