CPU overhead from malloc_usable_size

Using tools/benchmark.sh to run benchmarks shows that malloc_usable_size might account for between 10% and 20% of the regression between v6.0 and v6.28. Note that the regression from v6.0 to v6.28 is frequently in the range of 10% to 20%, 10% of 10% is 1% while 20% of 20% is 4%.

So disabling malloc_usable_size in v6.28 will make it between 1% and 4% faster. I am not suggesting that, but I am suggesting we learn whether the usage of it can be made more efficient.

The easy way to reproduce this is to run benchmark.sh for a workload that can be cached by RocksDB.