fix: persist SRM instance error on failed QoD/TI/EAM operations

When SRM reports a failed operation, the reason can sit only in
instances[].error with the top-level error left null. The QoD, TI and
EAM completion handlers saved only event.error, so operations.error
was null for every such failure.

Add completion_error() next to SRMOperationCompleted: it returns
event.error, or the first instance error when the operation failed
and event.error is missing. partially_completed operations still
keep operations.error null. Use it in all three handle_completed
methods.