Add a retry count to the subprocess task: re-run a failed child N times

Follow-up to the subprocess failure handling: a subprocess node now resumes its parent with __subprocess_failed__ when its child fails, so retry is already expressible in the workflow graph (route the failed branch back to the subprocess node, bounded by a counter). A first-class retry count would be a convenience on top.

Proposed solution: add an optional retry setting to the subprocess node. When the child fails, re-launch a fresh child up to N times (with an optional backoff), and only once the attempts are exhausted resume the parent with __subprocess_failed__.

• Same retry semantics as the node-advance retry, but a different mechanism: the node-advance retry re-runs a throwing advance within one transaction, whereas a subprocess retry re-launches a separate child instance.
• The attempt count is tracked per subprocess token; the failure-resume this builds on is the terminal once the budget is spent.