Slim Class-File method and class metadata to primitive access flags and descriptor (Java 24+)

[LordKay-sudo]

Follow-up to #37111 / #37112.

#37112 closed the largest part of the Class-File vs ASM heap gap on Java 24+ (84.5 MB to 16.7 MB in the reproducer). Its description notes that one difference is intentionally left for later:

ClassFileMethodMetadata holds AccessFlags and MethodTypeDesc objects and ClassFileAnnotationMetadata one AccessFlags, where SimpleMethodMetadata and SimpleAnnotationMetadata keep plain int and String values instead. [...] Once both are changed, the reproducer matches the ASM baseline of 12.3 MB and the test context mentioned above comes to 46.9 MB.

I'd like to finish that remaining part, but only if it's wanted.

Proposal

Store the same primitives the ASM implementations already keep, in the Class-File implementations:

• ClassFileMethodMetadata: int access + String descriptor (instead of AccessFlags + MethodTypeDesc)
• ClassFileAnnotationMetadata: int access (instead of AccessFlags)

Flag checks use java.lang.reflect.AccessFlag masks ((access & flag.mask()) != 0); toString() reconstructs a MethodTypeDesc from the stored descriptor only while building the string. Parameter type rendering from #36919 is preserved.

Not ASM

This deliberately avoids the reason #36978 was declined: it introduces no dependency on org.springframework.asm in the Class-File path. The only class-file imports used are java.lang.classfile.* and java.lang.reflect.AccessFlag.

Expected effect

From the reproducer numbers in #37112 (Temurin 25, spring-core 7.0.8):

	retained heap
SimpleMetadataReaderFactory (ASM baseline)	12.3 MB
ClassFileMetadataReaderFactory, current main (after #37112)	16.7 MB
ClassFileMetadataReaderFactory, with this change	12.3 MB (projected in #37112)

I validated the change structurally rather than re-running the full heap reproducer: on JDK 25 the spring-core test (ASM) and java24Test (Class-File, via the multi-release JAR) source sets are both green for the classreading and metadata tests, including the existing verifyToString / equals / hashCode contracts and a new java24Test asserting the fields are now primitives. Happy to run the reproducer for an independent figure if useful.

If you'd welcome this, I have a branch ready and can open a PR.