when the timestamp to start emitting updates from
is way in the past, CRDB will not emit checkpoints.

As a consequence, every update in the w.r.t to the
moment the changefeed was created will accumulate
in memory and OOM the process with a large enough
backlog.

The proposed solution is to compute one checkpoint
from the real-time stream of updates, and then use
that as the high-watermark for the backlog of
changes from the past. We need to emit all the updates
before we can emit a checkpoint, so downstream callers
would have to handle it accordingly.