perf(wave1): reduce Config caps, reorder Client/Monitor for cache

Wave 1 changes:
- Reduce CFG_MAX_KEYBINDS 256→128, CFG_MAX_STRLEN 256→64, CFG_MAX_ARGS 16→8
  → saves ~2.5 MB BSS from Config struct
- Add truncation warnings to config_load when keybind/button/scroll limit reached
- Reorder Client struct: HOT fields (type, mon, tags, floating flags, geom) in first
  cache line, COLD fields (wl_listeners) at end — saves cache misses in arrange()
- Reorder LayerSurface struct similarly
- Split Monitor into Monitor/MonitorCold: hot fields (~120B) stay inline, cold layout
  state lazily allocated on first dwindle/master use
- All cold-field accesses through m->cold-> indirection
- cleanupmon frees m->cold
- Zero-warning build with clang -Werror -Wpedantic

src/layout.h

@@ -26,6 +26,7 @@ void monocle(Monitor *m);
 /* helpers used across modules */
 [[nodiscard]] int current_tag_idx(Monitor *m);
 [[nodiscard]] const Layout *curlayout(Monitor *m);
+void ensure_cold(Monitor *m);
 
 /* dwindle tree helpers (exposed for swapdir) */
 [[nodiscard]] DwindleNode *dwindle_find_leaf(DwindleNode *n, Client *c);