perf(rocksdb): Add VoidNamespace fast-path and reusable read buffer

[nateab]

Summary

• Add VoidNamespace fast-path in SerializedCompositeKeyBuilder — short-circuits namespace serialization for the most common namespace type (non-windowed state) by writing the single byte directly instead of full serializer dispatch
• Add reusable value buffer for RocksDB read operations — replaces the allocating db.get(CF, key) API with db.get(CF, ReadOptions, key, keyOff, keyLen, valueBuf, valOff, valLen) which writes into a pre-allocated reusable buffer, eliminating one byte[] allocation per state read
• Applied to RocksDBValueState.value(), RocksDBMapState.get()/contains(), and AbstractRocksDBAppendingState.getInternal() (reducing/aggregating states)

Benchmark Results (vs master)

Benchmark	Master	Optimized	Change
ValueStateBenchmark.valueGet	793.7 ± 45.6 ops/ms	921.0 ± 63.9 ops/ms	+16.0%
MapStateBenchmark.mapGet	97.4 ± 2.3 ops/ms	102.9 ± 4.4 ops/ms	+5.6%

Note: These results include the cumulative effect of all optimizations on feature/rocksdb-zerocopy-key-serialization (write-path ByteBuffer API) plus this PR's read-path changes.

Design Decisions

• byte[] pre-allocated buffer instead of ByteBuffer for reads: The db.get(ColumnFamilyHandle, ReadOptions, ByteBuffer, ByteBuffer) variant requires direct ByteBuffers only (assertion in FRocksDB 8.10.0). The byte[] variant avoids this complexity while achieving the same allocation elimination.
• Buffer growth strategy: Start at 128 bytes, grow to exact size on first overflow. Amortized O(1) since values typically stabilize.
• Not applied to ListState.getInternal(): Uses ListDelimitedSerializer.deserializeList() which needs full byte[]. Follow-up optimization.
• Not applied to mergeNamespaces(): Loop with changing keys, low incremental gain.

Test plan

• Full RocksDB state backend test suite (839/839 passing, 0 failures)
• JMH benchmarks show improvement on read operations
• No regression on write operations