perf: Avoid constructor setup during resolve

[dcodeIO]

Resolving fields currently reads parent.ctor to install prototype defaults. Since ctor is a lazy getter, this also generates a runtime message constructor for every resolved type. That constructor setup then resolves the same fields again.

This PR separates reflection resolution from runtime constructor initialization:

• Field#resolve updates prototype defaults only for constructors that already exist, avoiding the lazy ctor getter.
• Type#ctor initializes prototype defaults when a constructor is actually created or registered.
• wrapper setup reuses the type's runtime ctor instead of creating a second one.
• additionally, recursive feature resolution is skipped when the feature tree is already clean.

On a synthetic benchmark with many cross-type references, this removes one redundant field-resolution pass and avoids generating constructors during schema loading, improving fromJSON by roughly 40-50% depending on schema size.

For 1500 types x 4 refs, the median local fromJSON time went from 9.150ms to 4.855ms. The relevant call counts changed from 12000 to 6000 Field#resolve calls and from 1500 generated constructors during load to 0.

So instead of rolling back eager resolution as proposed in #2157, this keeps the current eager semantics while reducing redundant work.

Follow-up to #2066 and #2068
Addresses #2123