Errors were not getting cleared from the queue between various OpenSSL
function calls. This was causing the error queue to contain misleading
information on subsequent calls to SSL_get_error.

One degenerate case was incorrect dispatching of errors after a failed
SSL_read. Conditions that simply should have lead to another call to
SSL_read were being treated as fatal. This was verified while stracing a
stud process and seeing read() return with EAGAIN and stud immediately
closing the socket fd and logging an SSL_read error.

Conflicts:
	stud.c