LPeg bug with nil captures

[Julien Desgats <julien+lua@...>]

Hello, 

This is definitely a bad day for LPeg :)

One of the fixes of LPeg 0.12.1 was related to nil captures, the fix
was to set an empty ktable as uservalue for Lua 5.2+. However, the
following code still have issues with LPeg 0.12.2:

 Lua 5.2.4  Copyright (C) 1994-2015 Lua.org, PUC-Rio
 > L=require 'lpeg'
 > =L.version()
 0.12.2
 > =(L.P"aaa" * L.Cc(nil)):match"aaa"
 lua: lapi.c:653: lua_rawgeti: Assertion `(((((t))->tt_) == (((5) | (1
 << 6))))) && "table expected"' failed.

This is the case only when the Lua VM is compiled with
-DLUA_USE_APICHECK, otherwise the behavior is not defined (sometimes
segfaults).

This is because the 'L.Cc(nil)' pattern do have an empty ktable, but
when merged with 'L.P"aaa"', the joinktables considers that neither of
source patterns have a ktable and keeps the nil uservalue of resulting
pattern. When matching is attempted, LPeg attempts to push the capture
by calling lua_rawgeti on that nil uservalue.

This issue does not occur with Lua 5.1 as all userdata have a table as
fenv.