The new SST iterators in #82799, e.g. NewPebbleSSTMemIterator(), are much slower than the existing NewMemSSTIterator(), around 80% or so. Some of this is due to CRDB-side overhead, in particular the use of pebbleIterator and VerifyingMVCCIterator, but the majority appears to be in Pebble via the new NewExternalIter() SST iterator. We need to optimize these.

This can be seen with BenchmarkSSTIterator, by changing NewMemSSTIterator() to NewPebbleMemSSTIterator() in runSSTIterator() on top of #82799.

Many of the optimizations that address #82559 and #83049 would likely apply here as well.

Jira issue: CRDB-16822

Epic CRDB-2624