The error types should be refactored

I have many issues with the various `Error` types we define and how `HttpsConnector` ultimately just uses `BoxError` for its `Service` error type, I'll try to summarize them here.

First, the `BoxError`, this makes it impossible to consume any more specific error type, as downcasting with the `Error` trait is always by reference.

Second, I keep confusing myself with `boring::Error` and `boring::ssl::Error`.

Third, the `boring::ssl::HandshakeError` is not fun to use for multiple reasons:
* the underlying I/O errors can be in two separate variants (`Failure` and `WouldBlock`);
* it stores a `MidHandshakeSslStream` even in the `Failure` variant, even though you are obviously not supposed to do anything anymore with that stream given the handshake failed;
* given that second bullet point, it also doesn't make much sense for the `MidHandshakeSslStream` struct to keep around the error that interrupted the handshake, as that was expected and you just want to resume it;
* it has a variant `SetupFailure` which feels out of place to me, shouldn't setup errors be completely contained in builders etc?

Fourth, `tokio_boring::HandshakeError` is as useful as `BoxError` given it doesn't let us access the `boring::ssl::HandshakeError` it wraps directly, so that's one more layer of hoops to go through to find, say, I/O errors.

Fifth, even if `tokio_boring::HandshakeError` let us access its inner `boring::ssl::HandshakeError`, that would still be a bit of a bother to use, as we *know* `tokio_boring` would never return a `WouldBlock` error but we would still need an arm for that in our code.