kv-client: use multiple sync.Map to reduce lock contention (#1439)

[ti-srebot]

cherry-pick #1439 to release-4.0
You can switch your code base to this Pull Request by using git-extras:

# In ticdc repo:
git pr https://github.com/pingcap/ticdc/pull/1476

After apply modifications, you can push your change to this PR via:

git push git@github.com:ti-srebot/ticdc.git pr/1476:release-4.0-355016c0ee9d

---

What problem does this PR solve?

part 3 of #1393, please review this PR after #1426 is merged

• In write less, read more scenario, sync.Map is faster than RWMutex and Mutex. Note this is concluded from some benchmarks, such as sync: RWMutex scales poorly with CPU count golang/go#17973, https://www.jianshu.com/p/cffffa914381, but maybe not quite suitable for our scene, we need more tests for this point, the most important thing here is to simulate the real world workload.
• Separate into multiple sync.Map can accelerate map access

What is changed and how it works?

• Use multiple sync.Maps for regionFeedState get/put operation.
• The bucket size of sync.Map pool is determined by cpu number, min threshold, and max threshold

Check List

Tests

• Unit test
• Integration test

Release note

• No release note

[ti-srebot]

/run-all-tests

[amyangfei]

/merge

[ti-srebot]

/run-all-tests