Reported by Arseny Vakhrushev.
Fix contributed by Peter Cawley.

This reverts the reversion that was done to merge with LuaJIT/LuaJIT.

Fixes #111.

feature: added new Lua API functions thread.exdata2 as well as C API funcs lua_getexdata2 and lua_setexdata2.

Actually move saved_jit_base to the end of global_State

I mentioned the change in the previous commit but somehow didn't
actually put it in.  Actually resolves #106.

Putting prng and saved_jit_base in the middle of the struct messed
with alignments on Armv7.  Avoid the problem by moving the prng to the
end so that the upstream hack of adding padding at the end works
correctly.

Fixes #106.

bugfix: compilation error "for’ loop initial declarations are only allowed in C99 mode" would happen with older gcc versions like 4.8.

This regression had appeared in commit 7c03b051.

The new hashing scheme in luajit is more sophisticated, with two
different hashing functions and a proper multilevel hash to account
for collisions.  The sparse hash function is expected to be fast and
the dense hash function is expected to be slower.  Accommodate the
CRC32 based hash implementation accordingly by splitting it into fast
and slow functions.

The hash_sparse crc32 implementation avoids using the slow 128 and
above path so that it can complement the speed requirements.  Based on
the benchmark it takes about 20% less time to compute hashes compared
to the upstream hash_sparse.  It regresses in the 64-128 byte range
and then again gains performance in the 128 byte and larger range.

The hash_dense implementation is consistently faster than the stock
hash_dense implementation, which scales linearly.  The standard
deviation reported by the tests are comparable for both sparse and
dense implementations.

This change splits out the PRNG so that the PRNG used by the JIT
compiler can be overridden without affecting the PRNG for the
allocator.

The API is changed to reflect the wider PRNG data, which is now 256
bits.  jit.prngstate returns an array with 8 32-bit integers and
similarly, it can be overridden with an array of up to 8 32-bit
integers.

Tests and README.md have been updated to reflect this change.

Reported by Arseny Vakhrushev.

Reported by Philipp Kutin.
Fix contributed by Peter Cawley.

This merge makes jit.prngstate unusable since the prng usage has
changed in LuaJIT/LuaJIT.

TODO:

- Make jit.prngstate usable with the new prng functionality
- Reinstate the string hashing optimisations

This partially reverts commit 6315a752
since lj_str_new no longer has the fast path.

Revert "optimize: lj_str_new: tests the full hash value before doing the full string comparison on hash collisions. thanks Shuxin Yang for the patch."

This reverts commit 00895099.

Revert "optimize: lj_str_new: uses randomized hash functions based on crc32 when -msse4.2 is specified."

This reverts commit 7923c634.

Revert "bugfix: fixed compatibility regression with MinGW gcc. this bug had appeared in commit 7923c634."

This reverts commit b982ebb5.

Revert "bugfix: FFI C parsers could not parse some C constructs like `__attribute((aligned(N)))` and `#pragma`."

This reverts commit 2d339277.

Reverts state back to LuaJIT/v2.1 way of parsing attributes.

This reverts commit 4589430e.

Revert "bugfix: fixed a clang warning in lj_str.c regarding unused str_fastcmp() when macro LUAJIT_USE_VALGRIND is defined."

This reverts commit 4e9ce0fb.

The fuction str_fastcmp has been dropped in LuaJIT/LuaJIT.  Also
remove note about this fix from README.md.

This reverts commit 3ad7e2eb.

This issue has already been fixed in LuaJIT/LuaJIT.

Reported by Jason Carr.

An unused guarded CONV int.num cannot be omitted in general.

Reported by Jason Carr.