xquartz: expose the Option layer as a real XKB level

[lummi-ch23]

Problem
On a non-US Mac keyboard layout, Option+key characters (the AltGr / level-3
layer, e.g. @ on Option+G, # on Option+3 on Swiss German) do not reach
XKB-native clients (GTK, Qt) over X11. xev reports the correct keysym, but GTK
apps get the base glyph (g instead of @).

Cause
DarwinBuildModifierMaps rewrites the Option key to XK_Mode_switch when
!XQuartzOptionSendsAlt, and the four glyphs per key are consumed as two
core-protocol groups. Core clients reach the Option layer as group 2 via
Mode_switch, but XKB-native clients resolve levels/groups through the XKB state,
where Mode_switch is a level shifter and the key types built from the width-4
core map are two-level. Level 3 is therefore unreachable and the app falls back to
the base glyph. The existing code even comments "Yes, this is ugly."

Fix
Map the Option key to XK_ISO_Level3_Shift on Mod5 and, after
XkbApplyMappingChange, install a genuine FOUR_LEVEL key type for each key that
carries an Option-layer glyph, laying the four glyphs out as one group of four
levels (base, shift, option, option+shift). XKB-native clients then reach the
Option layer as level 3, matching how a native Linux ch(de_mac) layout presents
it. Core-protocol behaviour is unchanged.

Testing
Built on Apple Silicon against the 2.8.6 SDK (-Dxquartz=true -Dsecure-rpc=false -Dxcsecurity=true), installed over X11.bin. Verified with a real keypress:
Option+G yields @ and Option+3 yields # in GTK apps (Intune portal,
xfce4-terminal); confirmed via xev and xkbcomp -xkb that the G key now has a
four-level type with @ at level 3 and Option on Mod5. ssh -X untrusted
forwarding works (SECURITY extension enabled).

[lummi-ch23]

Updated the branch: the first revision crashed the server (SIGSEGV in XkbCopyKeymap/_XkbCopyClientMap memmove during device-class copy on the next key event). Root cause: the new FOUR_LEVEL type slot was not zeroed before XkbResizeKeyType (XkbAllocClientMap does not clear a slot that fits existing size_types slack), leaving level_names a garbage pointer, and the key symbol map was hand-edited (width/offset) inconsistently. Fixed by zeroing + populating the type slot and assigning it via XkbChangeTypesOfKey, which resizes the symbol map and keeps width/offset/group_info consistent. Verified with a real keypress on Apple Silicon (2.8.6 SDK).