For a while, the cdc/crdb-chaos and cdc/sink-chaos roachtests have been
failing because an error that should be marked as retryable wasn't. As a
result of the discussion in cockroachdb#35974, I tried switching from a whitelist
(retryable error) to a blacklist (terminal error) in cockroachdb#36132, but on
reflection this doesn't seem like a great idea. We added a safety net to
prevent false negatives from retrying indefinitely but it was
immediately apparent that this meant we needed to tune the retry loop
parameters. Better is to just do the due diligence of investigating the
errors that should be retried and retrying them.

The commit is intended for backport into 19.1 once it's baked for a bit.

Closes cockroachdb#35974
Closes cockroachdb#36018
Closes cockroachdb#36019
Closes cockroachdb#36432

Release note (bug fix): `CHANGEFEED` now retry instead of erroring in
more situations

The `%v` and `%+v` were switched from what I intended in the line that
logs retryable errors and the one that handles non-retryable errors.

Release note: None