[arm] Flink documentation contains stale tuning and serialization claims

[MarcusKainth]

The Flink fairness review found documentation and validation rules that no longer describe the measured configuration. This issue tracks their correction; it does not claim a measured performance improvement.

Evidence at benchmark commit 9b4cab3:

• entrants/flink/README.md calculates 800,000 buffered/in-flight rows. The descriptor now uses 32 subtasks, 100,000 buffered rows and two 50,000-row requests: 3,200,000 buffered plus up to 3,200,000 in flight.
• The README says task off-heap memory is 128 MiB; config.yaml sets 512 MiB.
• The approximately 49-second drain and Docker Desktop explanation describe older experiments. The archived Linux/Graviton4 runs use 40 million messages and take approximately 988 seconds for Flink.
• The statements that POJO fields never use Kryo and that no serialization occurs need qualification. Chaining/object reuse avoid inter-operator copies; Avro decoding, ClickHouse encoding and checkpoint serialization remain. Nested field type extraction must be verified independently of the outer POJO type.
• Equal parallelism is only one chaining condition. Explicitly setting equal per-operator parallelism does not itself break chaining. Fewer consumers than partitions does not imply a universal half-rate penalty.
• The claims that larger heaps necessarily worsen GC and that the live set does not grow need run provenance. Current records show approximately 17.5 GiB post-GC occupancy against a similarly sized heap.
• flink_jvm_sizing_fits_its_declared_container enforces historical tuning choices as a minimum and a compressed-reference boundary as a maximum. These are not fairness requirements; legitimate smaller heaps and larger heaps must remain testable within the resource envelope.
• Spate's README still describes eight-shard aggregate concurrency and approximately 99.7% batch-cap fill. Its current RowBinary descriptor uses 32 shards, and current actual rows per INSERT need to replace historical batch-fill claims.

Acceptance:

• Correct configuration-derived statements and distinguish configured capacity from observed occupancy/batch size.
• Link empirical claims to exact run records, hardware, corpus and protocol; remove unsupported generalizations.
• Explain the difference between Flink POJO typing and the ClickHouse connector's typed mode.
• Document the actual checkpoint and connector serialization paths, effective chain and resolved field serializers.
• Replace historical heap restrictions with resource-envelope validation.
• Link the completed review and configuration changes here. Keep tuning records separate from published results.

This review is requested by the benchmark maintainer. All candidate changes preserve the contract's public-API requirement and five-second at-least-once durability.