[otel] End all recording spans on non-terminal invocations

[zhongkechen]

Summary

The OTel plugins intentionally drop references to some recording spans without ending them when a durable execution returns a non-terminal status.

Affected behavior:

• ExecutionOtelPlugin drops open operation spans on PENDING or RETRY.
• Both ExecutionOtelPlugin and InvocationOtelPlugin drop the execution-scoped Workflow span on non-terminal invocations.

OpenTelemetry requires every created span to be ended.

Evidence

• packages/aws-durable-execution-sdk-python-otel/src/aws_durable_execution_sdk_python_otel/execution_plugin.py, invocation-end handling around lines 275-337
• packages/aws-durable-execution-sdk-python-otel/src/aws_durable_execution_sdk_python_otel/invocation_plugin.py, workflow handling around lines 438-461
• packages/aws-durable-execution-sdk-python-otel/tests/test_execution_plugin.py::test_open_operation_span_not_exported_at_invocation_end explicitly asserts that an operation span is never ended
• packages/aws-durable-execution-sdk-python-otel/tests/test_invocation_plugin.py::test_workflow_span_not_exported_on_non_terminal codifies the same workflow behavior

A focused runtime probe retains the workflow span reference, invokes on_invocation_end with PENDING, and observes workflow.is_recording() is True after plugin state has been reset.

OpenTelemetry requirement

The Trace API states that every span must be ended and that failing to end spans may leak memory or other resources:

https://opentelemetry.io/docs/specs/otel/trace/api/#end

Impact

• Span processors can receive on_start without a corresponding on_end.
• Processor or exporter state may remain live across warm Lambda invocations.
• Re-created deterministic spans can trigger multiple starts for the same logical span ID.
• The behavior prevents a strict OpenTelemetry compliance claim.

Acceptance criteria

• Every recording span created by either OTel plugin is ended exactly once.
• Non-terminal workflow and suspended-operation behavior does not rely on abandoning a recording span.
• Cross-invocation workflow and operation correlation remains intact.
• Tests assert that retained span references are no longer recording after invocation cleanup.
• Tests cover PENDING, RETRY, terminal success, and terminal failure for both plugin variants.