refactor secio handshake

[Stebalien]

1. Respect context. Before, we didn't respect the context when writing/reading the final nonce.
2. Don't make assumptions about connection buffering. Read/write handshake messages in parallel.
3. Return messages from ReadMsg to the buffer pool.
4. Close the connection on timeout. We can't reuse it at this point as there may be an outstanding writer/reader (and we don't allow parallel reads/writes).
5. Make sure that ephemeral key generation actually succeeds.

Testing: The tests in go-libp2p-swarm and go-libp2p-conn still pass (and they're much more comprehensive than the ones here).

[whyrusleeping]

This LGTM, saves us from spawning new goroutines for each message send. Which I like.

[whyrusleeping]

I still would like to have some more tests here, but all things considered, I'll let it pass this time...

[Stebalien]

@whyrusleeping it only saves us a few (half - 1). I'm still using go routines to read and write a message at the same time.