isImplementedInClass leaks the losing sample instance on re-entrant or racing cache population

从 NativeScript/runtime/Metadata.mm 中的 MethodMeta::isImplementedInClass 以及指向此 issue 的注释开始；跟踪 sampleInstances 的填充、mutex 的使用和 alloc 的可重入行为。完成标准是在可重入或存在竞争的填充过程中消除被遗弃的 sample instances，同时不发生死锁、不安全释放或有意的永久保留。

Problem

MethodMeta::isImplementedInClass (NativeScript/runtime/Metadata.mm) keeps a process-wide cache of [klass alloc]-created sample instances used to answer respondsToSelector: for classes whose alloc returns a different class or which forward messages. The cache is populated like this:

1. [klass alloc] runs outside the mutex — deliberately, because alloc can trigger +initialize, which may run arbitrary code that re-enters this method (holding the lock would deadlock).
2. The lock is then taken and sampleInstances.emplace(klass, instance) inserts.

When the emplace loses — the re-entrant call already populated the entry for the same class, or another thread raced — the freshly allocated instance is abandoned: one leaked object per lost race (Instruments shows these as e.g. a leaked UIAlertView / NSURLSessionConfiguration attributed to isImplementedInClass). The count varies run to run since it is timing-dependent.

Why the obvious fixes don't work

• Releasing the loser is unsafe: the instance is alloc'd but never init'd, so -release runs -dealloc against zero-filled ivars of an arbitrary framework class, on whatever thread the probe ran on (it demonstrably runs on worker threads). Benign for most classes, but a -dealloc that does CFRelease/dispatch_release on a zero ivar traps, and UIKit teardown off the main thread is asserting territory.
• Holding the lock across alloc deadlocks via the +initialize re-entry described above.
• Parking losers in a static container merely converts the unreachable leak into intentional perpetual retention — it silences Instruments without reclaiming anything (tried and reverted in #458).

Potential solutions

• A re-entrancy-aware locking scheme (recursive mutex, or a reader/writer arrangement) so the populate path can be made atomic with respect to re-entrant probes without deadlocking through +initialize.
• Reuse already-alloc-ed objects: when a sample for that Class is requested (or something else calls [thatClass alloc] through the runtime), hand out / consume the cached instance instead of allocating another, so a losing instance gets used rather than abandoned.

The leak site carries a comment pointing at this issue.