Deadlock in the TLSEngine

[fcabestre]

When multiple streams are concurrently reading/writing a TLSSocket, there is a reproducible deadlock in the TLSEngine. This is illustrated by the following scastie.

Let me quote a quick analysis from @mpilquist on the FS2 Gitter room:

Definitely a bug in TLSEngine -- it occurs when there's a read that blocks for data followed by a write and the write starts the TLS handshake. The read gets the first response data from the handshake and also tries to start handshaking separately, but gets blocked on the handshake semaphore.