Update to tokio 0.3 and bump major version number

[olivia-fl]

Tokio 0.3 removed the PollEvented struct, which required restructuring things quite a bit in order to update tokio-libtls. In the process, I discovered some issues with the previous implementation. This update contains multiple breaking changes, so I bumped the version number to 2.0.0. Even without the API changes, increasing the required tokio version is a breaking change by itself, so I don't think there's a way to avoid this.

The main problem with the old implementation was that, for the implementations of AsyncRead and AsyncWrite, the current task was never scheduled to be woken in the future if IO would block. Instead, it simply returned Poll::Pending without doing anything. Handling this correctly requires registering interest in the socket being either readable or writable, depending on whether the libtls function returns TLS_WANT_POLLIN or TLS_WANT_POLLOUT. I ended up making a helper function poll_io to handle this logic, and removing the try_async_tls macro.

Side note: a thing that confused me a lot initially was that tls_read and tls_write can both return either TLS_WANT_POLLIN or TLS_WANT_POLLOUT. I initially assumed that tls_read would only need the socket to be readable and that tls_write would only need writable.

It turns out that poll_io is the exact same logic needed for processing the handshake in AsyncTls::poll, so I restructured that section of the code to use poll_io instead of the previous method which involved storing readiness inside of the Error variants. I'm not certain whether the old implementation of AsyncTls::poll ever worked correctly, but it was very confusing (for me at least) and polluted the public Error enum with variants that would never actually get returned from the function. After removing these extra variants, Error becomes redundant because it's only variant is TlsError. At this point I just removed it entirely. This is technically a breaking change, but I also don't think Error was ever returned from a function, so making it public in the first place was probably a mistake, and I don't imagine any code depending on tokio_libtls is matching against it. Since we're going to need breaking changes anyways in order to bump the tokio dependency version, we might as well remove Error. Another option would be to keep Error in the public api but mark it as deprecated since it's never constructed.

I also removed AsyncTlsStream entirely. Previously, AsyncTlsStream was useful for exposing the PollEvented methods publicly, but PollEvented no longer exists, and if you want to poll readable or writable readiness on the socket you can just use TcpStream::poll_read_ready and TcpStream::poll_write_ready. I'm not certain if there are any situations where users would want to do this by hand, but to allow this we could either make TlsStream::tcp public or write duplicate methods in TlsStream that just call the corresponding TcpStream methods. Either way the previous functionality of AsyncTlsStream is now a part of TlsStream and the old AsyncTlsStream type is unnecessary. Following similar reasoning to removing Error, I think this is going to have breaking changes anyway so we might as well remove it, but making type AsyncTlsStream = TlsStream and marking it as deprecated is also an option.

One last question is whether or not the tls and tcp members of TlsStream should be marked public. I'm not sure whether there are situations in which users would want direct access to these variables, but if so that's an easy change.

[olivia-fl]

So about whether or not tls and tcp should public: I was using this for my own project and ran into a situation that needed access to tls in order to look at certificates. I marked both of these fields as public, just in case anybody wants to poll the tcp stream by hand. One new question is whether or not TlsStream should implement Deref<Target=Tls>. I'm generally not in favor of that pattern, but it might be worth considering here.