Optimize dynamic schemas

[adwsingh]

What behavior changes?

No observable behavior change. Dynamic-schema document serde produces the same
output for the same input. This is purely internal: less allocation and faster
serialize/deserialize for StructDocument / ContentDocument, especially for
union-dense payloads (e.g. DynamoDB AttributeValue) and string/enum lists.

Why is this change needed?

Performance. The old path wrapped every leaf value in a second Document and
stored struct members in a LinkedHashMap keyed by member name. This added an
allocation per leaf plus map overhead per struct. The hot serialize call site
was also megamorphic, so C2 couldn't inline it.

How was this validated?

• Existing dynamic-schemas unit tests pass (./gradlew :dynamic-schemas:test).
• Serde benchmarks (the dynamic-client serde bench) used to measure the
  allocation / throughput wins.
• Protocol tests pass.

What should reviewers focus on?

• ContentDocument: now stores the raw unwrapped value (String/Number/Boolean/
  Instant/ByteBuffer/List/Map) instead of wrapping a Document. Look at
  serializeContents / serializeScalar (kept in sync by hand) and the
  string/enum list fast-path in serializeListContents.
• StructDocument + SchemaGuidedDocumentBuilder: struct members now live in a
  member-index-keyed Document[] instead of a LinkedHashMap; unions store the
  single set value raw (no ContentDocument wrapper). Reusable consumer objects
  replace per-call closures.
• SchemaDocument: the one case that still wraps a Document (untyped
  document-typed members).

---

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.