Fix concurrent publication of generated LuaObject metatables

[akeit0]

Concurrent first access to generated [LuaObject] metatables can expose an incomplete table even when callers use separate LuaState and object instances. The getter currently publishes the table before adding its metamethods, and competing initializers also mutate the shared field's current table.

Build each candidate locally, then publish it with Interlocked.CompareExchange after all metamethods are installed. Use Volatile.Read/Write for the shared property's access. Readers converge on the published instance, and initialization does not overwrite an existing non-null custom metatable. Explicit replacement and null-triggered reinitialization retain their existing behavior. Generated comments explain why publication occurs last.

Fixes #309.

Validation:

• Reproduced the original failure with eight threads, separate states/objects, and 500 synchronized initialization rounds. The regression checks all six generated metamethods, shared table identity, property writes, and Lua method/metamethod calls.
• Added replacement/reset coverage. Both new tests pass with the fix; the concurrent regression fails against main with missing metamethods and invalid userdata operations.
• Release runtime tests: 322 passed, excluding ExpectedFailure, matching CI. Source Generator tests: 1 passed.
• GitHub Actions lint and test checks passed for commit 816a8a6.
• Runtime builds and generated consumers (class, generic class, record) compile for netstandard2.1, net6.0, net8.0, and net10.0 with no warnings/errors. Formatting and diff checks pass.

Performance measured with BenchmarkDotNet 0.14.0, Windows x64/.NET 8.0.31, Release, two launches, five warmups, twelve measured iterations per launch. Both versions use the same main runtime; only the generator differs. The Lua workload performs 100,000 iterations of property reads and method calls and checks its result.

Workload	main mean	PR mean	Allocation per operation (both)
Cached metatable getter	1.305 ns	1.051 ns	0 B
Reset and initialize	49.252 ns	49.210 ns	848 B
Lua member access	15.345 ms	14.620 ms	70 B

No regression was observed in these measurements; the Lua workload took approximately 4.7% less time. Both cached-getter measurements have bimodal distributions, so their small absolute difference should not be treated as a reliable speedup. Unity Mono/IL2CPP execution was not measured.

The initialized getter takes no lock. Concurrent cold access can allocate losing candidate tables, which are discarded. This fix makes initial construction/publication safe; it does not synchronize subsequent mutations of a shared metatable or concurrent use of one LuaState.

Temporary benchmark sources, consumer projects, and logs remain outside the repository. This PR is left unmerged for the maintainer's final decision.