RFC: add benchmark for Compact() (and possible performance improvements)

[mitake]

Currently bbolt doesn't have benchmarks for Compact() function (which will be used for etcd's defrag: etcd-io/etcd#15470).

Performance of compaction is important. It's one of the top contributors for 5xx of k8s API server. So I think it's valuable to add some benchmarks of the compaction logic for understanding its performance properties and possible improvements. There might be some interesting parameters and possible ideas which are valuable to check, like:

• how does FillPercent affect write throughput and disk efficiency?
• if the compaction process can skip fdatasync() during intermediate tx, how will this affect throughput?
• can bulk-loading of B+tree improve throughput?

Of course corruption issues should be higher priority but I think it's valuable to have some benchmarks.

cc @ahrtr @serathius @ptabor @chaochn47 @cenkalti @geetasg