support write slow down on a single shard.

[coocood]

We should limit the maximum number of L0 tables, otherwise, the shard size will keep growing and never have the chance to split.

[coocood]

We can maintain a golang.org/x/time/rate.Limiter in each Shard to limit the write throughput.

The applier doesn't limit the write because that would block other region's write.

The Shard adjusts the limit based on the number of L0 tables and MemTable size, and it provides an API to be used by the MVCCStore.Prewrite method to limit the write throughput, add limit here doesn't block any other threads.