X11: General cleanup

[francesca64]

This is almost entirely internal changes, and as usual, doesn't actually fix any problems people have complained about.

• XSetInputFocus can't be called before the window is visible. This was previously handled by looping (with a sleep) and querying for the window's state until it was visible. Now we use XIfEvent, which blocks until we receive VisibilityNotify. Note that this can't be replaced with an XSync (I tried).
• We now call XSync at the end of window creation and check for errors, assuring that broken windows are never returned. When creating invisible windows, this is the only time the output buffer is flushed during the entire window creation process (AFAIK). For visible windows, XIfEvent will generally flush, but window creation has overall been reduced to the minimum number of flushes.
• check_errors().expect() has been a common pattern throughout the backend, but it seems that people (myself included) didn't make a distinction between using it after synchronous requests and asynchronous requests. Now we only use it after async requests if we flush first, though this still isn't correct (since the request likely hasn't been processed yet). The only real solution (besides forcing a sync every time) is to handle asynchronous errors asynchronously. For future work, I plan on adding logging, though I don't plan on actually handling those errors; that's more of something to hope for in the hypothetical async/await XCB paradise.
• We now flush whenever it makes sense to. util::Flusher was added to force contributors to be aware of the output buffer.
• Window::get_position, Window::get_inner_position, Window::get_inner_size, and Window::get_outer_size previously all required several round-trips. On my machine, it took an average of around 80µs. They've now been reduced to one round-trip each, which reduces my measurement to 16µs. This was accomplished simply by caching the frame extents, which are expensive to calculate (due to various queries and heuristics), but change infrequently and predictably. I still recommend that application developers use these methods sparingly and generally prefer storing the values from Resized/Moved, as that's zero overhead.
• The above change enabled me to change the Moved event to supply window positions, rather than client area positions. Additionally, we no longer generate Moved for real (as in, not synthetic) ConfigureNotify events. Real ConfigureNotify events contain positions relative to the parent window, which are typically constant and useless. Since that position would be completely different from the root-relative positions supplied by synthetic ConfigureNotify events (which are the vast majority of them), that meant real ConfigureNotify events would always be detected as the position having changed, so the resultant Moved was multiple levels of misleading. In practice, this meant a garbage Moved would be sent every time the window was resized; now a resize has to actually change the window's position to be accompanied by Moved.
• Every time we processed an XI_Enter event, we would leak 4 bytes via util::query_pointer (XIQueryPointer). XIButtonState contains a dynamically-allocated mask field which we weren't freeing. As this event occurs with fairly high frequency, long-running applications could easily accumulate substantial leaks. util::PointerState::drop now takes care of this.
• The util module has been split up into several sub-modules, as it was getting rather lengthy. This accounts for a significant part of this diff, unfortunately.
• Atoms are now cached. Xlib caches them too, so XInternAtom wouldn't typically be a round-trip anyway, but the added complexity is negligible.
• Switched from std::sync::Mutex to parking_lot::Mutex (within this backend). There appears to be no downside to this, but if anyone finds one, this would be easy to revert.
• The WM name and supported hints are now global to the application, and are updated upon ReparentNotify, which should detect when the WM was replaced (assuming a reparenting WM was involved, that is). Previously, these values were per-window and would never update, meaning replacing the WM could potentially lead to (admittedly very minor) problems.
• The result of Window2::create_empty_cursor will now only be used if it actually succeeds.
• Window2::load_cursor no longer re-allocates the cursor name.
• util::lookup_utf8 previously allocated a 16-byte buffer on the heap. Now it allocates a 1024-byte buffer on the stack, and falls back to dynamic allocation if the buffer is too small. This base buffer size is admittedly gratuitous, but less so if you're using IME.
• with_c_str was finally removed.
• Added util::Format enum to help prevent goofs when dealing with format arguments.
• util::get_property, something I added way back in my first winit PR, only calculated offsets correctly for util::Format::Char. This was concealed by the accomodating buffer size, as it would be very rare for the offset to be needed; however, testing with a buffer size of 1, util::Format::Long would read from the same offset multiple times, and util::Format::Short would miss data. This function now works correctly for all formats, relying on the simple fact that the offset increases by the buffer size on each iteration. We also account for the extra byte that XGetWindowProperty allocates at the end of the buffer, and copy data from the buffer instead of moving it and taking ownership of the pointer.
• Drag and drop now reliably works in release mode. This is presumably related to the util::get_property changes.
• util::change_property now exists, which should make it easier to add features in the future.
• The EventsLoop device map is no longer in a mutex.
• XConnection now implements Debug.
• Valgrind no longer complains about anything related to winit (with either the system allocator or jemalloc, though "not having valgrind complain about jemalloc" isn't something to strive for).