Indexing and repository import no longer slow themselves down writing their own progress log

Re-indexing a content collection, or importing a repository, writes a line of progress for everything it touches. Recording one of those lines means rewriting the whole progress log, so a run that touched a thousand files rewrote a thousand-line document a thousand times. The work of reporting the job grew faster than the job itself — a long run spent more of the server's time on its own log than on the indexing, and everything else sharing that server slowed down with it.

The per-file detail is now collected and written once per collection, which makes the cost grow in step with the number of files instead of with its square. Nothing is lost: every file is still named in the log, in order, and still lands before the run reports finished. A long walk also emits a short "N files processed" line as it goes, so you can still see it is alive.

Repository import got the same treatment for files it cannot parse. Those used to be reported one write at a time — and, because import parses several files at once, those writes competed with each other on the same record. They are now reported together, in a single entry that still names every file that was skipped and why.