With MMBIN instructions, there are fewer opcodes that need to update
'L->top', so that change does not seem to pay for the increased
complexity.

Instead of updating 'L->top' in every place that may call a
metamethod, the metamethod functions themselves (luaT_trybinTM and
luaT_callorderTM) correct the top. (When calling metamethods from
the C API, however, the callers must preserve 'L->top'.)

In 'lua_settop', avoid calling 'luaF_close' when increasing the stack
or when the function has no to-be-closed variables.

When leaving a scope, the new stack top should be set only after
closing any upvalue, to avoid manipulating values in an "invalid"
part of the stack.

Checks of the form '1 <= x && x <= M' were rewritten in the form
'(unsigned)x - 1 < (unsigned)M', which is usually more efficient.
(Other similar checks have similar translations.) Although
some compilers do these optimizations, that does not happen
for all compilers or all cases.

- The warning functions get an extra parameter that tells whether
message is to be continued (instead of using end-of-lines as a signal).

- The user data for the warning function is a regular value, instead
of a writable slot inside the Lua state.

Several small improvements (code style, warnings, comments, more tests),
in particular:

- 'lua_topointer' extended to handle strings
- raises an error in 'string.format("%10q")' ('%q' with modifiers)
- in the manual for 'string.format', the term "option" replaced by
  "conversion specifier" (the term used by the C standard)

After a major bad collection (one that collects too few objects),
next collection will be major again. In that case, avoid switching
back to generational mode (as it will have to switch again to
incremental to do next major collection).

Errors in finalizers (__gc metamethods) are never propagated.
Instead, they generate a warning.

The warning system is just a way for Lua to emit warnings, messages
to the programmer that do not interfere with the running program.