Summary:
This patch optimized Get() code paths by avoiding malloc of iterators. Iterator creation is moved to mem table rep implementations, where a callback is called when any key is found. This is the same practice as what we do in (SST) table readers.

db_bench result for readrandom following a writeseq, with no compression, single thread and tmpfs, we see throughput improved to 144958 from 139027, about 3%.

Test Plan: make all check

Reviewers: dhruba, haobo, igor

Reviewed By: haobo

CC: leveldb, yhchiang

Differential Revision: https://reviews.facebook.net/D14685