Fix segfault when input events arrive during initialization

[Yiin]

Summary

Fixes the segfault reported in #181 caused by pointer/keyboard events arriving before cursor surfaces and xkb state are initialized.

• Move cursor_surface creation and cursor theme loading before the second wl_display_roundtrip(), which is the first point where the compositor dispatches input events via newly-created layer surfaces
• Add defensive NULL checks in pointer_handle_enter, keyboard_handle_key, and keyboard_handle_modifiers as a safety net

Root cause

During initialization, seat listeners are registered in the first roundtrip (via handle_global → create_seat), but cursor_surface is only created after the second roundtrip. When the compositor sends a pointer_enter event during that second roundtrip (triggered by the layer surfaces committing), pointer_handle_enter dereferences a NULL cursor_surface inside wl_surface_set_buffer_scale() or wp_cursor_shape_manager_v1_get_pointer().

The same class of bug affects keyboard_handle_key and keyboard_handle_modifiers, which can dereference NULL xkb_state if a keyboard event arrives before the keymap event.

Reproduction

The race is timing-dependent but becomes nearly deterministic with many outputs (e.g. 45+ headless virtual monitors on Hyprland), as the roundtrip takes longer to enumerate all outputs, widening the window for events to arrive before initialization completes. Using wayfreeze --after-freeze-cmd "slurp ..." also exacerbates it since slurp immediately inherits pointer focus.

Crash backtrace:

#0  wl_proxy_get_version (libwayland-client.so.0)
#1  pointer_handle_enter (slurp)
...
#10 wl_display_roundtrip_queue (libwayland-client.so.0)
#11 main (slurp)

Fixes #181