[Feature]: Add retryable SerDes errors and targeted retry policies

## What would you like?

Add a retryable SerDes error classification and targeted retry policies for transient serialization/deserialization failures.

Currently `safeSerialize()` and `safeDeserialize()` treat any thrown error as terminal and terminate the invocation. Async SerDes implementations that use S3 Files, EFS, DynamoDB, KMS, or another external service need a way to classify transient failures and retry only that transformation.

## Possible Implementation

- Introduce a public `RetryableSerdesError` (following the repository's naming conventions) that retains the original cause.
- Add a `retrySerdes()` decorator, and make the same retry behavior available to individual stages if a composable pipeline is introduced.
- Retry only the retryable error type; all other SerDes failures should follow the existing immediate termination behavior.
- Accept a bounded retry policy with maximum attempts and optional asynchronous backoff/jitter, reusing existing retry-strategy concepts where appropriate.
- Pass the same value/data and `SerdesContext` to every attempt.
- After exhaustion, preserve the cause and terminate with the existing `SERDES_FAILED` reason.
- Ensure retries wrap only the selected SerDes transformation rather than restarting the durable operation.
- Add fake-timer tests for success after retry, exhaustion, non-retryable failures, serialize/deserialize parity, and context stability.

## Is this a breaking change?

No.

## Does this require an RFC?

No.

## Additional Context

Python already exposes `RetryableSerDesError`, while the Java SDK work in aws/aws-durable-execution-sdk-java#648 adds a targeted retry decorator for individual SerDes stages. TypeScript currently has neither a retryable SerDes classification nor a targeted retry wrapper.

Start by locating safeSerialize(), safeDeserialize(), SerdesContext, and the existing retry-strategy concepts in the TypeScript SDK. Review how SerDes failures produce the SERDES_FAILED reason, then add fake-timer coverage for retry success, exhaustion, non-retryable failures, serialize/deserialize parity, and context stability. Done means only the selected transformation retries while other failures retain existing termination behavior.

Written by the indexing model from the issue text.