mlua-sys: bind LuaJIT's native luaL_traceback instead of the compat53 port

[ross]

Under the luajit and luajit52 features, luaL_traceback resolves to the lua-compat-5.3 port
in mlua-sys/src/lua51/compat.rs. LuaJIT exports its own, and the port's cost grows with the
size of the script's global tables. mlua builds a traceback for every callback error and every
error that escapes to Rust, including errors a script catches with pcall, so ordinary error
handling pays it.

This PR binds LuaJIT's luaL_traceback and compiles the port only for Lua 5.1.

Why the port is slow

To name a frame, compat53_pushfuncname calls compat53_findfield over _G to depth 2: every
global and every field of every global table. On main it tries the global name first for every
frame, so every traceback walks _G, and a frame that isn't a global function (the main chunk,
for one) walks all of it. Lua 5.4 searches package.loaded instead. LuaJIT searches nothing: it
prints the call-site name, or the C function's address.

Measurements

Per caught error, min to max over 24 fresh VMs, each holding a 20,000-entry global table. Release
build, LuaJIT 2.1 from luajit-src, x86-64 Linux. rust.refuse is a Rust function that returns
Err.

mlua main (mlua-sys 0.13.0):

Case	Port	Native luaL_traceback
pcall(error, "x") (control, no callback)	1.0-2.7 µs	1.1-1.6 µs
pcall(rust.refuse)	242-925 µs	3.2-4.6 µs
pcall(function() rust.refuse() end)	515-1,039 µs	3.6-5.7 µs

mlua 0.9.9 (mlua-sys 0.6.8):

Case	Port	Native luaL_traceback
pcall(error, "x") (control, no callback)	1.7-2.6 µs	1.7-2.5 µs
pcall(rust.refuse)	6.6-596 µs	3.9-4.8 µs
An error from a Rust __newindex metamethod	5.2-8.6 µs	4.7-5.3 µs

The control row's spread comes from the host's mixed core types.

Change

• lua51/lauxlib.rs: declare luaL_traceback in the extern block under
  #[cfg(feature = "luajit")]. The signature matches the Lua 5.2-5.4 declarations, so no caller
  changes.
• lua51/compat.rs: put #[cfg(not(feature = "luajit"))] on the port's luaL_traceback, the
  helpers only it uses (compat53_countlevels, compat53_findfield,
  compat53_pushglobalfuncname, compat53_pushfuncname, COMPAT53_LEVELS1,
  COMPAT53_LEVELS2), and the CStr and mem imports.

Compatibility

• LuaJIT versions. LuaJIT added luaL_traceback before 2.0.0 (commit "From Lua 5.2: Add
  luaL_traceback().", 2012-10-01, v2.0 branch), so every version build/find_normal.rs accepts
  from pkg-config (2.0.4 to 2.2) exports it, as does the luajit-src 2.1 vendored build.

• Stack use. LuaJIT grows the stack as it pushes and concatenates every 15 pieces, so
  LUA_TRACEBACK_STACK still covers it.

• Traceback text. Output follows LuaJIT's format, the one debug.traceback already uses under
  LuaJIT. Code that parses traceback text for a qualified name gets the call-site name instead:

  Frame	Port (main)	LuaJIT
  Unnamed C function	[C]: in function 'rust.refuse' if found in _G, else [C]: in ?	[C]: at 0x55cd55a80520
  Unnamed LuaJIT builtin	[C]: in function 'NAME' if found in _G, else [C]: in ?	[builtin#N]:
  Named C function	[C]: in function 'table.sort'	[C]: in function 'sort'
  Named Lua function	in local 'f', in method 'm'	in function 'f'

Testing

• main: cargo test --features luajit,vendored,async,serde,macros passes before and after (358
  passed, 8 ignored). cargo clippy on mlua-sys is clean under luajit,vendored,
  luajit52,vendored, and lua51,vendored.
• mlua-sys 0.6.8 with mlua 0.9.9: the same change through [patch.crates-io] passes a downstream
  application's 323 tests, including ones that assert on traceback content.

🤖 Generated with Claude Code

[ross]

Ran into this when i was seeing huge variability in test run times, a specific test case with that was sometimes a few seconds, others 1.5 minutes.

/cc ross/logit#499