Support perfect retriever evaluation in Reader nodes

[tstadel]

According to #1719 in pipeline.eval(), we want to evaluate reader nodes besides the usual pipeline evaluation as if there was a perfect retriever sitting in front of them. Needs to be done:

• extend pipeline.eval() to receive a parameter that enables perfect retriever evaluation
• extend Reader.run() to receive the parameter enabling perfect retriever evaluation and the labels
• implement a second predict pass that operates on perfect documents from the labels
• store result of second predict pass in result dict
• ensure that second predict result is stored in _debug output
• extract second predict result in pipeline.eval() and write it to dataframe using the node_input column from Support more than one answer/document output per node in pipeline.eval() #1850 to distinguish it from default results
• extend pipeline.print_eval_report to show metrics of multiple node_inputs if provided