[FEATURE][Java] Pure-Java GraphAr reader/writer implementation

## Describe the enhancement requested

Track the production-driven pure-Java data-access implementation as one umbrella. The product fork develops it immediately and independently; this Apache issue organizes only the small upstreamable slices.

Component(s): Java

## Proposed layering

`graphar-info` -> storage API -> IO request/capability API -> Parquet backend -> GraphAr chunk/layout reader -> ordered-by-source `neighbors` and `scanEdges` -> writer/validator.

## Non-goals for the first vertical

No JNI/FastFFI, mandatory Arrow representation, general query AST, Gremlin, graph algorithms, ORC/CSV, unordered layouts, cloud storage, or integrations.

## Acceptance for the first vertical

A Java reader resolves `vertexId -> offsets -> minimum Parquet range -> destinations` for `ordered_by_source`, exposes streaming edge scans, and preserves the dependency boundary. Writer, other layouts, and adapters are separate small PRs after the reader is proven.

## Links

Architecture umbrella: #756. Compatibility/TCK umbrella: #944. Open a storage, IO, Parquet, reader, or writer subissue only when its independently verified fork slice is ready to be proposed upstream.

Start by reading the architecture umbrella in #756 and the compatibility/TCK umbrella in #944, then review the proposed layers from graphar-info through the Parquet backend and ordered_by_source reader. No file, test, or entry point is named; done means a separately verified slice can resolve vertex IDs to destinations, stream edge scans, and preserve the dependency boundary.

Written by the indexing model from the issue text.