TLS clients don't (currently) add a certificate, so there's no need
to delete it.

Use cryptDestroySession() instead.

With this, it's not strictly necessary to clear cert_list in ssl_sync()
when the certificate changes, but it's still a good idea to prevent
unusable memory from being held onto.

warning: value computed is not used

Now each lock has an easily understandable purpose, and covers the
lowest possible amount of code.

Use new rwlock for ssl certificate.

See merge request !386

Instead of tight loops with 1ms Sleep()s in them, add events for
zero readers and zero writers that we can wait for instead.

Unfortunately, since Events aren't interlocked with a critical section
like condition variables are with mutexes in pthreads, we can't
rely on this for race-free code, so for read locks, we still may
spin under write pressure.

Various bits of broken no longer broken.

Since we can't bool anyway, may as well.

It would be nice to use TLS for the recusrion counter, but on Windows,
it looks like TLS in threads created after the object is created is
uninitialized, so we would need some kind of thread start hook thingie
which just seems like too much work.

Instead, rely on the old standby single-linked list to track.

It looks like Windows actually uses the name for matching existing
events, so if two events have the same name, they end up being the
same object.

Recursive read locks, non-recursive write locks.

Currently the "wrapper" is just the POSIX API.
We may not need the pthread_rwlock_timed??lock() functions, and even
if we do, we like want a different wrapper that takes a ms timeout
instead of an absolute time.

On Win32, the name param is an LPCSTR which apparently is Microsoft
talk for const char *.

In C we're allowed to do whatever we want, but in C++, this matters
more.

Just running git commit --amend doesn't do -a it seems. :)

This should fix a long-standing issue where someone could connect to
the SSH port and do nothing, which would prevent other incoming
terminal sessions from being accepted until it times out.

Unfortunately, this means that Synchronet can't send any data until
authentication is completed, which means useful messages about why
you're being disconnected (ie: "Sorry, all terminal nodes are in use or otherwise unavailable.")
as well as usless information nobody ever cares about (ie: The IP
you're connecting from, that it is resolving your hostname, etc).
can no longer be sent to the user.

Reported by Max (WESTLINE)...

Introduced in commit 75c5c9e5, since we are now calling the set_state()
callback periodically, and in sbbsctrl, it sets/resets the pause button
checked (depressed) state in response to this callback (should it?), this was
resetting the pause button checked/depressed state peridically (at the
configured sem file check interval, default 2 seconds).

The fix: if the new server state is the same as the previously set server
state, don't call the set_state() callback. I considered this as a potential
optimization when I first wrote that commit, but didn't see any obvious
downside to calling the set_state() callback redundantly. Now that I see a
downside, let's implement this saving/checking of the state to avoid
redundant "set" calls.

... with a socket descriptor value of -1.