Deprecate aggregate_logging_outputs API (use reduce_metrics instead)

[myleott]

In 1e324a5 we introduced fairseq.metrics, which allows metrics to be logged and aggregated based on their context. For example:

with metrics.aggregate("train"):
    for step, batch in enumerate(epoch):
        with metrics.aggregate("train_inner") as agg:
            loss = get_loss(batch)
            metrics.log_scalar("loss", loss)
            if step % log_interval == 0:
                # print the average loss over this log_interval
                print(agg.get_smoothed_value("loss"))
                agg.reset()
# print the average loss over the whole epoch
print(metrics.get_smoothed_value("train", "loss"))

This interface allows one to log metrics from anywhere, without having to pass the values up the call stack, resulting in a much simpler train.py.

To make the transition smoother, we introduce FairseqTask.reduce_metrics and FairseqCriterion.reduce_metrics to replace the old aggregate_logging_outputs method, although we provide backward compatibility for old Tasks/Criterions.

[facebook-github-bot]

This pull request was exported from Phabricator. Differential Revision: D19292402

[facebook-github-bot]

@myleott merged this pull request in 8679339.