Wire RichParser node visitors via DI factory instead of private property hack

[TomasVotruba]

Follow-up to #8208 — same goal, no PrivatesAccessor on PHPStan internals.

Before

PHPStanContainerMemento::removeRichVisitors() read the private RichParser::$nodeVisitors property, filtered it, and wrote a new collection back. Every PHPStan internal change breaks it (that is what #8208 had to repair, and the 2.2.6 crash before it).

$nodeVisitorsCollection = $privatesAccessor->getPrivateProperty($richParser, 'nodeVisitors');
// ... filter ...
$privatesAccessor->setPrivateProperty($richParser, 'nodeVisitors', new DirectExtensionsCollection($nodeVisitorsToKeep));

After

A small factory service builds the parser through the public constructor, and the config picks the visitors:

    currentPhpVersionRichParser:
        class: PHPStan\Parser\RichParser
        factory: @rectorRichParserFactory::create()
        arguments!: []
        autowired: no

    rectorRichParserFactory:
        class: Rector\DependencyInjection\PHPStan\RichParserFactory
        arguments:
            parser: @currentPhpVersionPhpParser
            nodeVisitors:
                - @PHPStan\Parser\AnonymousClassVisitor
                - @PHPStan\Parser\ArrayMapArgVisitor
        autowired: no

PHPStanContainerMemento is deleted.

Why a factory, and not just a nodeVisitors: argument

A plain argument in the config is silently dropped. PHPStan\DependencyInjection\AutowiredExtensionsExtension::beforeCompile() walks all definitions and overwrites every #[AutowiredExtensions] parameter with the lazy container-backed collection — but it skips definitions whose creator is not the class itself, i.e. factory-created ones.

Do we still need the filtering at all?

Checked: with all 23 visitors enabled the whole test suite passes and bin/rector process --dry-run over this repo produces identical output — the #9492 node-replacement bug is fixed upstream. The filter is still worth keeping for speed, running all visitors on every parsed file costs roughly 10% more CPU on a full run of this repo (~510s -> ~550-600s user time, 30s -> 37s wall).

Regression test for #9492 (PHPStanPrinterTest) stays, now without the manual visitor-removal call, since the container wiring does it.