storage: randomized testing of error propagation

Read errors are unlikely in modern local filesystems, and it's suspected that the error propagation codepaths in pkg/storage's EngineIterator and MVCCIterator are infrequently exercised. When external sstables are in use (ie, during online restore), ephemeral errors are more likely to occur. If these errors are not propagated, they can result in silently incorrect results. Randomized testing of the pkg/storage iterators' error propagation would help shake out any instances where we do not propagate ephemeral errors.

#123478 is related.

Jira issue: CRDB-46796