Handling error types when the decision is crucial for correctness/performance

[Lorak-mmk]

There are multiple places in the driver where we inspect the error, and perform some actions based on what it is.
Those places include:

• retry policies
• speculative execution
• lbp latency measure
• USE KEYSPACE results handling

In #1083 all those places gained a lint that forces precise handling of those errors (no wildcards). What is missing is an informed decision about each error on how to handle .
In order to not delay the PR any more, let's continue the discussion here.

[Lorak-mmk]

One more place: handling errors in schema agreement check. After #1473 we now always repeat the check (until timeout runs out). We should investigate if there are types of errors for which we can fail immmediately (like we did before for all errors).