Fix memory leak: stamp $raw in the endpoint layer instead of a Valinor converter

[loevgaard]

Fixes #7

Problem

Long-running processes (e.g. Symfony Messenger workers) leaked ~70KB per mapped object, eventually hitting the PHP memory_limit. Root cause is upstream CuyZ/Valinor#800: with a converter registered, ValueConverterNodeBuilder::unstack() calls Reflection::function() per converted value node, and Closure::fromCallable() on a non-\Closure callable (our invokable RawStamper) creates a fresh closure each call — so the memoized ReflectionFunctions accumulate forever in a static cache GC cannot touch.

One nuance found during verification: the leak requires a non-Closure callable. Closure::fromCallable() on an existing \Closure returns the same instance, so plain closure converters get a stable cache key and don't leak (they still pay per-node reflection overhead).

Fix

• Remove ->registerConverter(new RawStamper()) from Client::configureMapperBuilder() and delete src/Mapper/RawStamper.php.
• Stamp $raw in ResourceEndpoint::mapResource() — the single funnel behind getOne()/createOne()/updateOne()/getPage()/paginate() — via a new recursive walker (src/Response/RawStamper.php).
• The walker preserves the converter's polymorphic behavior exactly: the top-level Resource gets the full decoded body; every nested Resource gets its slice, matched by property name and list position — including through non-Resource DTOs (Order::$customer, Collection<T> items, Line::$product, the recursive Customer → CustomerContact → Customer graph). All pre-existing $raw tests pass unchanged, and Payload::fromResponse() on embedded resources keeps working.
• Bonus: $raw stamping is now builder-independent, so it works even when a consumer passes a bare MapperBuilder. The do-not-register-converters constraint is documented on configureMapperBuilder(), in the README production section, and in CLAUDE.md.
• The PHPStan purity-suppressor include stays (the issue suggested it could be dropped with the converter): PHPStan ignores @param docblocks on closures entirely, so the remaining pure-in-practice registrations (Payload null-stripper, Identifier::fromReference()) can't be proven pure. The comment in phpstan.neon.dist now explains the real reason.

Tests

• tests/Client/MapperMemoryLeakTest.php — regression guard: 50 collection pages must grow memory_get_usage() by < 1MB. Memory-based rather than reflecting into Valinor's private static cache, so it survives Valinor refactors and catches any reintroduced per-mapping growth. Verified to discriminate: temporarily re-adding an invokable no-op converter fails it with ~33MB growth.
• tests/Response/RawStamperTest.php — 10 direct unit tests covering every walker branch (nested resources, descent through non-Resource DTOs, recursive graph, missing/non-array/misaligned slices, literal raw key). 100% MSI on the walker.

Verification

composer phpunit (244 tests), composer analyse (PHPStan level max), composer check-style, vendor/bin/rector --dry-run, and vendor/bin/infection (MSI +19pp over threshold) all pass.