Better cleanup of output buckets after failed datums

[emk]

This is a follow-on to fixing #33.

From the source:

            // Remove `OutputFile` records for this datum, so we can upload the
            // same output files again.
            //
            // TODO: Unfortunately, there's an issue here. It takes one of two
            // forms:
            //
            // 1. Workers use deterministic file names. In this case, we
            //    _should_ be fine, because we'll just overwrite any files we
            //    did manage to upload.
            // 2. Workers use random filenames. Here, there are two subcases: a.
            //    We have successfully created an `OutputFile` record. b. We
            //    have yet to create an `OutputFile` record.
            //
            // We need to fix (2b) by pre-creating all our `OutputFile` records
            // _before_ uploading, and then updating them later to show that the
            // output succeeded. Which them into case (2a). And then we can fix (2a)
            // by deleting any S3/GCS files corresponding to `OutputFile::uri`.