The `luaL_testudata` and `luaL_checkudata` functions correctly treat the
absence of a metatable for the type to check as the type not matching,
so there is no need to have the type's metatable in memory until an
object using it exists. Therefore we can defer the metatable's creation
until the first time an object of that type is created.
Saves some 4500 bytes of Lua heap RAM, at least if no such objects are
created. The created metatables are shared between all scripts, so there
is no increase in the total RAM usage if all objects are used. However,
there is some increased risk of heap fragmentation and unexpected
out-of-memory errors on rarely used code paths.
It would be nice to construct the object in the new utility function,
but it's not possible in C++ to take the address of a constructor.
Instead passing in a lambda is ugly, hardly any more efficient, and
screws up constructors which take parameters.
Also optimizes by using the generated userdata creators where possible,
and convincing the compiler to omit unnecessary checks.
The latter returns a null pointer (and sets the string size to zero) if
the conversion fails, which the following code neglects to check,
potentially leading to null pointer dereferences.
Instead use the former, which throws a helpful type error if it fails
and doesn't expose the code to null pointers.
The Lua stack is guaranteed to have at least LUA_MINSTACK (default 20)
slots upon entry to C. Check to see if we might need more than that
minimum and only in that case call the function to check and resize the
stack. In virtually all cases the check can then be optimized away.
Additionally remove the redundant "Out of stack" message. Lua already
says "stack overflow" and a null message is valid.
Saves ~330B.
