fix: fix bugs for considering eviction in parallel and others

[lanlou1554]

1. get_data_from_cache and put_data_to_cache aren't atomic. Think about one thread executes get_data_from_cache very early, and it is swapped out, another thread complete put_data_from_cache for the same key, and this thread wakes up, try to execute put_data_from_cache, it will poll from s3, update cache again, but there is no need. So I still want to use status_of_keys to track all the keys in the mem and disk replacer, and if it is completed, then threads can directly return, which is the most efficient way to handle it. So I fix bugs for these situations.
2. when a key in status_of_keys is set to completed, then it will forever stay in status_of_keys even if it is evicted, so some later threads cannot find data if it is evicted. We have to remove the key from status_of_keys when it is evicted.
3. During eviction from mem to disk, we need to use status_of_keys as a huge lock to lock it, otherwise eviction & write conflicts may happen
4. when main insertion progress fails, it should notify waiters
5. if no one is waiting, then we don't need to grab the lock for replacers to pin (pin_count is 0)
6. for disk_store read_data, the key to unpin replacer is wrong

TODO: when mem_replacer wants to evict A from mem to disk, is there any possibility that the status of A is incompleted? (Currently I just use continue to aviod eviction to disk and write to disk conflicts...)

[lanlou1554]

If what I design is correct, then we can close #49