[Feature]: Deduplicate SerDes deserialization within an invocation

## What would you like?

Deduplicate repeated SerDes deserialization within a single Lambda invocation.

Context-dependent SerDes implementations may perform filesystem or network I/O. Repeated access to the same checkpointed result or exception currently invokes the configured SerDes again, even when the serialized input and durable payload identity are unchanged.

## Possible Implementation

- Add an invocation-scoped deserialization cache around SDK-managed SerDes calls.
- Key entries by SerDes instance identity, durable execution ARN, operation/entity identity, payload kind when available, and a digest of the serialized input.
- Cache successful values, including `None`, but never cache failures.
- Bound completed entries with an LRU or weak-reference policy so large deserialized values are not retained indefinitely.
- If SDK-managed child/concurrency threads can request the same value concurrently, share one in-flight result rather than issuing duplicate external reads.
- Clear all entries at the end of the Lambda invocation; do not use a process-global cache across replays.
- Add tests for cache hits, distinct contexts/data, `None`, failures, eviction, and concurrent callers.

## Is this a breaking change?

No.

## Does this require an RFC?

No.

## Additional Context

This is primarily valuable for filesystem and other external-storage SerDes implementations. The Java SDK work in aws/aws-durable-execution-sdk-java#648 uses an invocation-scoped bounded cache and in-flight deduplication.

Empieza localizando las llamadas SerDes gestionadas por el SDK y el ciclo de vida de la invocación de Lambda donde podría establecerse y limpiarse una caché limitada al ámbito de la invocación. Añade pruebas que cubran aciertos de caché, contextos y datos distintos, valores None, fallos, eviction y llamadores concurrentes; el trabajo estará terminado cuando las solicitudes coincidentes repetidas se dedupliquen sin almacenar en caché los fallos ni cruzar los límites de la invocación.

Escrito por el modelo de indexación a partir del texto del issue.