perf(mapper): cache per-class reflection plans during hydration

[osbre]

Hydrating rows into models now caches reflection metadata, primary key scans, and resolved mappers to avoid redundant per-row work.

Improvements

• Per-Class Reflection Plans: Caches reflection metadata, property types, and parent-child relations by class rather than recalculating them per row.
• Cached Primary Key: Memoizes the primary key property lookup on model inspectors to prevent thousands of redundant scans per nested eager-loading query.
• Singleton Mapper Cache: Replaces a factory-level cache that never hit (a fresh factory is built per map() call) with a container-scoped #[Singleton], so the cache lives and dies with its container - no cross-container leaks. It implements Resettable, so workers calling resetContainer() between requests re-resolve mappers instead of holding ones built from unregistered dependencies. ObjectFactory::__construct() takes a third argument as a result.

Performance Impact

Benchmark	Before	After	Change
Mapper stage 2,000-row SQLite select	92.6ms	46.7ms	-49.6%
End-to-end Model::all()	114ms	68ms	-40.3%
Eager-loading 50 users with nested relations	1,517ms	551ms	-63.7%
map() in a hydration loop Per call, steady state	0.372ms	0.016ms	-95.7%

[github-actions]

Benchmark Results

Comparison of perf/mapper-hydration against 3.x (b1ae60ea63210569329e04ea911b471d7d9fd9d0).

Open to see the benchmark results

No benchmark changes above ±5%.

_{Generated by phpbench against commit 981b7a0}

[xHeaven]

I've added 2 failing tests, please fix those regressions.