refactor(spanner): buffer chunks and add non-blocking isDataAvailable to GrpcValueIterator - #14699

… to GrpcValueIterator

Refactor GrpcValueIterator and GrpcResultSet to support non-blocking data availability checks and bounded chunk buffering. This is Step 2 of the asynchronous stream refactoring plan, allowing reactive and async result set consumers to verify whether a full row is available in memory without blocking the calling thread on network I/O.

Summary of changes:
- Bounded chunk buffering: GrpcValueIterator maintains an internal queue of PartialResultSet chunks, buffering only as many chunks as needed to assemble at least one complete row (including handling chunked string and list merges across multiple chunks).
- Non-blocking isDataAvailable(): Implemented isDataAvailable() on GrpcResultSet and GrpcValueIterator. Polling uses non-blocking tryLock() coordination so checking data availability never blocks behind concurrent worker threads reading from the stream.
- Zero-allocation hot path: Optimizes row availability checks to avoid heap allocations on the steady-state path when rows are already available in the current chunk.

Bug fixes and safety improvements included:
- Fix getMetadata() idempotency: Calling getMetadata() multiple times or mid-stream now returns cached metadata immediately without erroneously advancing the stream or discarding chunks.
- Prevent state resurrection on cancellation: If close() is invoked while a background stream.next() call is in flight, late-arriving chunks are safely discarded rather than repopulating state on a closed iterator.
- Immediate surfacing of malformed metadata: Streams returning an initial chunk without valid metadata or row types immediately signal isDataAvailable() so the internal error is thrown rather than hanging the caller.
- Clean up unused error field in GrpcResultSet.
- Added comprehensive unit tests covering all chunked merge permutations, 0-column queries, error paths, and deterministic concurrent access.