This commit addresses issue open-telemetry#1832.

The way `NET_PEER_IP` and `NET_PEER_PORT` are retrieved raises a `ValueError`
when gRPC connections are handled via Unix sockets.

```py
ip, port = (
    context.peer().split(",")[0].split(":", 1)[1].rsplit(":", 1)
)
```

When using an address like `unix:///tmp/grpc.sock` the value of `context.peer()` is `"unix:"`.
Substituting that in the function above...

```py
ip, port = "unix:".split(",")[0].split(":", 1)[1].rsplit(":", 1)
ip, port = ["unix:"][0].split(":", 1)[1].rsplit(":", 1)
ip, port = "unix:".split(":", 1)[1].rsplit(":", 1)
ip, port = ["unix", ""][1].rsplit(":", 1)
ip, port = "".rsplit(":", 1)
ip, port = [""]  # ValueError
```

I "addressed" the issue by guarding the retrieval of `net.peer.*` values under
an `if` statement that checks if we are using a Unix socket.

I extended the `server_interceptor` tests to run against TCP and Unix socket configurations.

---

**Open Questions**

- [ ] The socket tests will fail on Windows. Is there a way to annotate that?
- [ ] Are there other span values we should be setting for the unix socket?