The allocation of a userdata for the state of the warn system can
cause a panic if it fails; 'luaL_ref' also can fail. This commit
re-implements the warn system so that it does not need an explicit
state. Instead, the system uses different functions to represent
the different states.

Hashes for long strings are computed only when they are used as keys
in a table, not a too common case. And, in that case, it is to easy to
force collisions changing only the characters which are not part of the
hash.

The stack size is derived from 'stack_last', when needed. Moreover,
the handling of stack sizes is more consistent, always excluding the
extra space except when allocating/deallocating the array.

Test uses an extra bit in 'marked' to mark all elements in gray lists
and then check against elements colored gray.

When entering generational mode, all objects are old. So, the only
objects that need to be in a gray list are threads, which can be
assigned without barriers. Changes in anything else (e.g., weak
tables) will trigger barriers that, if needed, will add the object
to a gray list.

The fields 'old' and 'finobjold' were renamed 'old1' and 'finobjold1',
respectively, to make clearer the main ages of their elements.

It helps to have this function available for debugging.

Main thread must be non yieldable even at "level 0" (bare API), outside
the 'pcall' from 'lua.c'.

To allow their use in memory tests, some functions in 'ltests.c'
should never allocate memory. To avoid this allocation, the
library registers the strings used for status codes, and keeps
the variable '_WARN' always defined (with false instead of nil).

Since commit ca6fe7449a74, userdata with uservalues can be gray
and can belong to gray lists ('gclist').