fix(jira): stop the changelog convertor hiding unconverted items forever (#9042)

The incremental filter on the changelog convertor used created_at:

  _tool_jira_issue_changelog_items.created_at >= ?

created_at is stamped when the row is first inserted and never moves again.
So a changelog item that was collected during one window but not converted
in that window can never be selected by any later incremental run -- the
timestamp it is filtered on is permanently in the past. Nothing errors; the
rows simply stay in the tool layer.

That matches the report in #8834: a partial, silent shortfall in
issue_changelogs from the same sync run, persisting across runs, varying by
project. It also explains why the reporter's check looked clean -- they
inspected _devlake_collector_latest_state, which is the collector's state,
not the convertor's.

Switched to updated_at, which the extractor's upsert refreshes
(OnConflict{UpdateAll: true}), so a re-collected item is reconsidered. Of
the 36 convertors in the code base, this was the only one filtering on
created_at; the other 35 already use updated_at.

This does not widen the board filter, so it does not carry the cost klesh
raised against moving board_id into the join: no board task converts
anything outside its own board.

Separately, the board filter is now reported rather than silent. After a
successful conversion the subtask counts collected changelog items whose
issue is not on this board and logs the number with the reason. That is one
aggregate query per board task, and it turns an unexplained shortfall into
a logged figure. Diagnostic failures are logged, never propagated.

Tests: an e2e dataflow test driving the convertor with a changelog for an
issue on no board, asserting it is excluded, that nothing is attached to
issue id 0, and that in-scope changelogs still convert; plus a regression
guard on the filter column, since reverting it is a one-token change that
silently restores permanent data loss.