Replace pool with single gRPC client

[palkan]

Improve gRPC performance and stability by getting rid of the connections pool in favor of the built-in streaming and multiplexing provided by the grpc package.

Benchmarks show that using a single client result in less errors and slightly better performance than using a pool.

Also, retry mechanism has been also improved to handle exhausted and unavailable errors differently: when we receive ResourceExhausted we try to re-run the request almost immediately. The reason is that we assume that the concurrency settings for the client and the server are similar (though not equal, see below), thus, there could not be real exhaustion but the lag in server implementation (it looks like the worker is returned to the pool after the request completed, so there is a little gap).

For the same reason we set (and recommend setting) the default concurrency for RPC client to be a slightly less than the number of workers in the server (the corresponding defaults are 28 < 30).