LWS wasn't giving us the usual connection closed callbacks due to our early hijack of the connection. Instead, delete the buffer when we unregister the client's poll FD.

resolves #28

I'd been assuming that the OS always copied our USB transfer buffer at submit-time. Turns out that this is true for Linux but not for Windows or Mac. This caused tearing when fcserver received a new frame while previous frame(s) were still queued in the USB host controller.

Fixed this by copying to a per-transfer buffer on Windows and Mac. Linux is unaffected. We avoid performing an additional copy there.

If we issue a short write, libwebsocket handles it kinda ungracefully and we have no idea when to close the socket. Instead of letting libwebsocket copy our content into a temporary buffer, we can just give it smaller pieces.

This lets us uniquely identify a particular device connection instance.

This keeps the size of the fcserver binary down.

This was causing deadlocks all over on Windows. Now things seem to be
running smoothly!

This is necessary for URBs which synchronously complete. This is common on
the win32 backend.