[wolf4ood]

What this PR changes/adds

• DataPlaneSignalingClient classifies failures:

  • a signal that could not be delivered (timeout, I/O error), a 5xx response of the data plane, or an authorization that could not be obtained → ERROR_RETRY;
  • a 4xx → FATAL_ERROR.

  The transfer process state machine retries ERROR_RETRY with its backoff and retry limit (edc.transfer.send.retry.*) for every signal (prepare, start, suspend, resume, terminate, started, completed).

• DataPlaneSignalingFlowController documents the known limitations of a retried prepare/start.

Why it does that

Every signaling failure was final (StatusResult.fatalError), so a slow or briefly unavailable data plane terminated transfers instead of delaying them (#6177).

Further notes

Known limitations, documented in the javadoc and not addressed here:

• The HTTP client still re-sends a start/prepare whose response timed out. If the data plane created the flow anyway, it rejects the duplicate with a 4xx, and the transfer fails. The same applies to a retry by the state machine. Recovering from a lost synchronous response would need idempotent start/prepare in the Data Plane Signaling specification.
• A retried prepare/start selects the data plane again, so with several matching data planes it may go to another one.

Who will sponsor this feature?

Please @-mention the committer that will sponsor your feature.

Linked Issue(s)

Closes #6234

Please be sure to take a look at the contributing guidelines and our etiquette for pull requests.