Improvements for failover of Distributed queries

[Enmk]

I hereby agree to the terms of the CLA available at: https://yandex.ru/legal/cla/?lang=en

For changelog. Remove if this is non-significant change.

Category (leave one):

• Improvement

Short description (up to few sentences):
Improvements for failover of Distributed queries. Shorten recovery time, also it is now configurable and can be seen in system.clusters.
...

Detailed description (optional):

• Added a limit on how many errors can replica accumulate
• Decreased default error halving time to 60 seconds
• Made both configurable via settings: replica_error_max_count and replica_error_decrease_period
• Showing errors count and estimated recovery time for each replica in system.clusters

Example

SELECT *
FROM system.clusters

┌─cluster──────┬─shard_num─┬─shard_weight─┬─replica_num─┬─host_name─┬─host_address─┬─port─┬─is_local─┬─user────┬─default_database─┬─errors_count─┬─estimated_recovery_time─┐
│ test_cluster │         1 │            1 │           1 │ localhost │ 127.0.0.1    │ 9000 │        1 │ default │                  │            6 │                     166 │
│ test_cluster │         1 │            1 │           2 │ localhost │ 127.0.0.1    │ 9001 │        1 │ default │                  │            0 │                       0 │
│ test_cluster │         1 │            1 │           3 │ localhost │ 127.0.0.1    │ 9002 │        1 │ default │                  │            6 │                     166 │
└──────────────┴───────────┴──────────────┴─────────────┴───────────┴──────────────┴──────┴──────────┴─────────┴──────────────────┴──────────────┴─────────────────────────┘

• errors_count is number of times this host tried to reach replica but failed
• estimated_recovery_time how many seconds are left until replica error count is zeroed and replica is considered to be back to normal

Please note that errors_count is updated once per query to the cluster, but estimated_recovery_time is recalculated on-demand. So there could be a case of non-zero errors_count and zero estimated_recovery_time, that means that next query will zero errors count and will try to use replica as if it had no errors.
...

Fixes #5317

[alexey-milovidov]

Let's merge with master, because we have fixes for integration and performance tests.

[akuzm]

The docs should be updated to describe the new settings and the new fields in system.clusters.

[Enmk]

Docs updated, settings renamed.