perf: avoid shared pointer contention in manifest and Avro decode

[gripleaf]

Purpose

Linked issue: #240

Reduce shared pointer contention in concurrent manifest reads and Avro decoding hot loops.

This change:

• Caches StructArray::fields() once per batch in ManifestFile::ReadBucketEntries() and
  ObjectsFile<T>::Read(), avoiding repeated copies of shared_ptr<Array> for every row.
• Caches immutable Arrow builder metadata in AvroDirectDecoder::DecodeContext, including:
  • Arrow type ID.
  • Timestamp time unit.
• Uses the cached metadata for integer and timestamp decoding instead of calling
  ArrayBuilder::type() for every value.
• Clears cached builder metadata when SetReadSchema() replaces the builder tree, preventing stale
  entries if builder addresses are reused.
• Adds Int16 decoding and builder metadata lifecycle coverage.

Tests

• Built paimon-avro-format-test successfully.
• Ran the complete Avro unit test suite: 68 tests passed.
• Ran pre-commit run --files for all changed files: passed.
• Ran git diff --check: passed.
• paimon-core-test could not be fully linked because of an unrelated existing Parquet/Arrow API
  mismatch in page_filtered_row_group_reader.cpp: the current Arrow GetColumn overload expects
  const std::vector<int>&, while the existing source passes
  std::shared_ptr<std::unordered_set<int>>.

API and Format

No public API, storage format, or protocol changes.

Documentation

No new user-facing feature or documentation change.

Generative AI tooling

Generated-by: OpenAI Codex (GPT-5)

[gripleaf]

StructArray shared_ptr lock contention:

Avro decoding hot loops:

[lxy-9602]

LGTM. @zjw1111 , PTAL.

[lxy-9602]

Could you please create a corresponding issue for this PR? For example, if there have recently been several PRs focused on improving scan performance, it would be helpful to open a dedicated issue describing that effort. Thank you!

[zjw1111]

+1