fix: strengthen type tokens extraction

[romm]

This commit refactors the way tokens are extracted from a raw type string, by using a better approach that consists in splitting first into tokens, and then detecting texts (aka string values).

This aims to fix edge cases like the following example, where the $bar annotation would previously have confused the lexer, leading to the type of $foo being used for $bar as well.

final class SomeClass
{
    /**
      * @param non-empty-string $foo Some description containing $bar
      *      which is the next parameter name
      */
     public function __construct(
         public string $foo,
         public int $bar,
     ) {}
}

Fixes #494