The trigger was this line in the recent change to guru.dat:
Unrealized potential. :-(

That open paren was parsed as the beginning of a new guru expression. A very
long guru expression that overflowed the stack variable 'str' here.

Add setwinsize and setwinpos
Move blit from rect into gdi_thread
Add beep support
Add (non-functional) settitle support

When running Windows 11 in a VirtualBox VM, Sleep(1) appears to take
much more than 1ms to return.  This causes large delays drawing menus
using WIN_DYN (as SyncTERM does).

If WIN_DYN is set, just use kbhit() with no delay.  Leave the delay
there for !WIN_DYN though.

This should generally speed up SyncTERM menu navigation as there
was a 50ms delay for each menu on the screen.

unresolved external symbol _try_gdi_init referenced in function initciolib

Function is dependent on WITH_GDI, not _WIN32

There was a lot of weird things going on to avoid locking vstatlock
inside of screenlock.  Simply this by copying needed vstat stuff
while we hold the lock.

Also, get SyncTERM to stop displaying (null) as the mode.

Video is very slow in the menus, and the input appears to repeat
forever when connected (but *not* in the menus), but it's something
at least.

You can see some of these is in a msg_hdr.field_list, but these you won't:
- SMB_GROUP (never added by anything)
- RFC822MSGID (already represented by the id property)
- RFC822REPLYID (already represented by the reply_id property)

This will allow a "true" Windows native version, and hopefully be
at least as good as SDL mode so I can relegate that to a backup.

DDMsgReader: In MsgHdrPropIsKludgeLine(), removed checks for X-FTN-* fields, since those shouldn't exist

See merge request !289

Some Linux systems ship with a reentrant ncurses library that does
not have an ESCDELAY variable.  For now, simply don't allow control
over ESCDELAY on the systems since we also currently have a global,
and so can't use set_escdelay() as intended.

Also, in ripper, stub out all the stuff that's used when HAS_VSTAT
is not defined.  This makes most of the code not compiled when it's
useless anyway.

This works around an issue on macOS and Linux where a select() or
poll() in thread X won't return when close() is called on the same
socket in thread Y.

We also set the descriptor/handle/whatever to an invalid value
while waiting for io threads to terminate so that a close() in one
thread won't cause other threads to possibly interact with a new
descriptor created after the close().  This isn't actually something
that can happen in SyncTERM, but it's worth protecting against.

Finally, don't call close() until after all io threads have
terminated.  This prevents the descriptor from becoming available
to other syscalls while threads are closing, and prevents operations
on an invalid descriptor.

This is still not safe on platforms where a descriptor value write
is not atomic, but I don't think I support any platform like that.

Reported as #212 in github, and 75 on SourceForge.

RIP initialization was trying to obtain a lock that only exists when
using a bitmap console.  Don't allow RIP to be enabled, and don't
perform the operations that require the lock when the CONIO_OPT_SET_PIXEL
ciolib option is not set.

DDMsgReader: In MsgHdrPropIsKludgeLine(), removed checks for X-FTN-* fields, since those shouldn't exist

DDMsgReader: When viewing message headers, now shows all available message header values

See merge request !288

Hopefully nothing still builds conio as single-threaded anymore.

Pass the struct vmem_cell to bitmap_draw_one_char() rather than
having it rummage in vmem for it since the caller invariable has
it available.

Fixes issue on Linux where the window size is constantly updated,
and window rabidly steals focus.

A bit more cleanup here would be good, but I think this is pushed
down enough finally.

We only really ever treat it as a single lock, so we may as well
simplify this.

This lock has devolved into a wrapper lock for vstatlock.

Use pthread_once() for init_mouse() instead of a global
Fix some lock issues in cb_drawrect()

update_pixels should only be set when at least one pixel has
actually been changed... setting it when pixels have not been
changed causes unneeded screen updates.

Previously, a font change asserted (incorrectly) that the pixels in
the screen had been modified, so the change would occur at the end
of the next blink cycle when alt fonts are checked.  Now it
requests a full redraw from vmem when a font is changed, which will
trigger a redraw the next time the blinker thread triggers (~5ms).

More precalculation in bitmap_draw_one_char()
Optimize loop-initialization of buffers

This is about as good as it can get without drastic data structure
overhaul.  Biggest win with overhaul would be with eliminating
memmove() for scrolling and copying into scrollback. memmove()
accounts for a high percentage of CPU utilization right now.

When fast scrolling, memmove() takes most of the CPU by far... it
is likely actually worth having each line be a buffer of its own
and just moving the pointers around.  To help with this optimization
in the future, move the screen memmove() bit into a separate function.

Based on qwk.ssjs

I tested this using:
wget --auth-no-challenge --post-file=file.ext --http-user=username --http-password=passwd vert.synchro.net/upload.ssjs?filename=file.ext\?desc=description-text

This currently requires an "uploads" directory to be configured by the sysop.

I'm not sure if there's a better/more-standard way for the posted filename or
content to be included in the request. This is just a sort of proof-of-concept
as a solution for issue #554. Perhaps an ecbwebv4 page displays a prompt to the
user for their filename and it redirects to this page. Not sure how that'll
work.