now we require OpenResty's LuaJIT branch to work properly on ARM64 since
we make use of the thread.exdata Lua API.

On architectures other than ARM64, the thread.exdata API also saves the
per-request global env table and closure objects for each light thread,
which gives a nice ~10% speedup for the simplest Lua handler location
loaded by wrk over HTTP 1.1.

We now use proper userdata instead of lightuserdata for the
ngx.shared.DICT Lua objects.

we also mask off the bits higher than 47-bit of lightuserdata keys based
on C static variable addresses.

We also implemented the pure C API for the ndk.* API (to be used in
lua-resty-core).

Thanks Dejiang Zhu and Zexuan Luo for the development work of this patch.

Signed-off-by: Yichun Zhang (agentzh) <agentzh@gmail.com>

…ecursively.