fix(core): keep TypeDef root kind consistent for factory custom serializers

[Pigsy-Monk]

What does this PR do?

Fixes #4021.

A fresh reader can infer a struct kind for a plain class whose configured
SerializerFactory actually supplies an extension serializer. Native TypeDef
validation then rejects bytes written by an identically configured instance.

Resolve empty extension roots through the existing ClassResolver.getTypeInfo
path before validating their kind. The resolver retains the actual serializer
and type ID, including when metadata previously created a provisional TypeInfo.
This requires no additional cache or factory probe.

Keep each serializer's root definition on TypeInfo, independently of the class
field schemas used by metadata probes and ObjectStream layers. Local-header
matching checks an available root definition before the field schema, without
creating a serializer just to probe for a match. This preserves cold registered
struct and ObjectStream layer matches and prevents a cached field schema from
replacing an extension's empty root definition.

Tests

Expanded SerializerFactoryTest covers the original polymorphic custom serializer,
metadata-first and typed-root reads, read-then-write reuse, enum and recursive
struct factories, factory failure recovery, registration, and both codegen modes.
It also verifies that each custom serializer is created once per resolver.

• Full fory-core suite on JDK 21: 2,252 passed, 58 skipped, no failures or errors.
• Includes cold native ID/name registration and typed/dynamic reads, verifying
  that exact-local schemas do not consume the remote-schema allowance, plus
  existing ObjectStream layer, registration, and schema-consistent meta-share cases.
• Local cross-language peer tests were not run; the change preserves the native wire format.
• Spotless and Checkstyle passed.