Revert "When monitoring idle connections, only call eof on tcp socket"

[quinnj]

Reverts #912

The problem here is that when we go to reuse connections later, the ssl layer hasn't had time (or been told by monitor_idle_connections) to process encrypted data and potential close notify records. This can be bad if you, for example, do a bunch of simultaneous requests which build up a bunch of MbedTLS.SSLContext connections in the connection pool, then go to try and do a request later and all the connections are actually closed, even though they haven't been properly closed at the ssl layer.

[vtjnash]

Yeah, I don't think that PR was right. Violating the abstraction of the underlying socket by examining its internal state is likely to lead to problems.