iterator: allow NextPrefix from exhausted position

[RaduBerinde]

We don't need to support NextPrefix from an exhausted position, but we
want to implement predictable semantics for the metamorphic tests.

The iterator could be exhausted in either direction; in both cases,
returning whatever Next returns is correct.

Fixes #3597

Also see #3593 (review)

[cockroach-teamcity]

This change is 

[RaduBerinde]

Thinking about this some more - maybe the problem was that even with the previous code, most of the time NextPrefix returns a key, after a SeekLT returned nil. That would suggest that we don't (always) set the iterValidityState to IterExhausted when SeekGT returns nil, should we?

[RaduBerinde]

This diff also seems to fix the failure:

--- a/iterator.go
+++ b/iterator.go
@@ -1589,8 +1589,8 @@ func (i *Iterator) SeekLTWithLimit(key []byte, limit []byte) IterValidityState {
        if seekInternalIter {
                i.iterKV = i.iter.SeekLT(key, base.SeekLTFlagsNone)
                i.stats.ReverseSeekCount[InternalIterCall]++
-               if err := i.iter.Error(); err != nil {
-                       i.err = err
+               if i.iterKV != nil {
+                       i.err = errors.CombineErrors(i.err, i.iter.Error())
                        i.iterValidityState = IterExhausted
                        return i.iterValidityState
                }

[RaduBerinde]

Closing in favor of #3604 (we posted at the same time).