Replace upvalue-based shell dispatch with OP_LUSH opcode

Add a dedicated OP_LUSH A B instruction that loads shell C functions
directly from the registry, eliminating the need for upvalue allocation
in mainfunc(), upvalue population in lua_load(), and the fragile
LUSH_NUM_UPVALS heuristic. Every chunk no longer carries 4 extra
upvalues.

lvm.c

@@ -1939,6 +1939,19 @@ void luaV_execute (lua_State *L, CallInfo *ci) {
         Protect(luaT_getvarargs(L, ci, ra, n));
         vmbreak;
       }
+      vmcase(OP_LUSH) {
+        StkId ra = RA(i);
+        int idx = GETARG_B(i);
+        Table *regt = hvalue(&G(L)->l_registry);
+        TValue lushtv;
+        lu_byte tag = luaH_getint(regt, LUA_RIDX_LUSH, &lushtv);
+        if (novariant(tag) == LUA_TTABLE) {
+          TValue func;
+          luaH_getint(hvalue(&lushtv), idx + 1, &func);  /* 1-based */
+          setobj2s(L, ra, &func);
+        }
+        vmbreak;
+      }
       vmcase(OP_GETVARG) {
         StkId ra = RA(i);
         TValue *rc = vRC(i);