While scanning a non-unique secondary index of a vinyl space, we skip
a tuple if there's no match for it in the primary index because it means
that the tuple was overwritten by a newer statement but the secondary
index DELETE was deferred until compaction. We still create a cache
chain link in this case so that we skip the overwritten tuple when
we scan the secondary index next time. We do it even if the overwriting
statement is unconfirmed (not yet written to WAL), which is actually
incorrect. The problem is that if the overwriting statement is rolled
back on a WAL write error, the chain created in the secondary index
cache won't be broken because nothing was actually written to the
secondary index (DELETE was deferred) therefore there's nothing to
roll back therefore there's nothing to invalidate the cache by.
As a result, the tuple that the user attempted to delete but failed
due to a WAL error will become inaccessible via the secondary index
because of the stale cache chain.

Fix this issue by disabling creation of a cache chain link in case
the read iterator skipped an unconfirmed deferred DELETE statement.

Closes tarantool#11140

NO_DOC=bug fix