Improving some internal error-handling

[dominicbarnes]

The goal of this PR is to set us up to improve visibility into errors being returned by kafka-go by wrapping those errors before bubbling them up throughout the call stack. The hope is that by making each error into a clear chain, it should be much easier to troubleshoot by narrowing in on the most relevant branch of code that is failing.

To start towards this more lofty goal, the first step (and the main focus of this PR) is to replace "strict" error-checking (eg: err == io.EOF) with more modern/flexible checks (eg: errors.Is(err, io.EOF)). This will allows us to freely wrap errors internally while not changing any error-handling logic at higher levels. Without this, blindly wrapping errors could result in subtle bugs as the older methods of checking errors would no longer behave the same way.

While I did include some error-wrapping changes, the changes are pretty focused on the error-checking instead, which is already a decent size PR. We can use subsequent PRs to refactor other components (eg: reader, writer) and catch any other error-checking places I may have missed.

[dominicbarnes]

I had made some changes to conn.go, but it looks like the x/net/nettest package explicitly expects an unwrapped net.Error which causes the test to fail when a fmt.Errorf is used.

It's possible this is just an expectation of a net.Conn interface/contract we can't/shouldn't change, so we may have to tread more carefully there.