[TLX] Add tlx.threadfence() and tlx.threadfence_system()

[dshi7]

Triton TBE backward uses a multi-CTA cooperative pattern where multiple thread blocks atomically accumulate partial gradients, then the last block applies the optimizer update. This requires a GPU-scope memory fence between gradient writes and counter decrement — equivalent to CUDA's __threadfence(). TLX previously had no GPU/system-scope fence.

The new ops lower through TTNG_ThreadfenceOp to LLVM::FenceOp with the appropriate syncscope, which the NVPTX backend emits as fence.acq_rel.gpu / fence.acq_rel.sys.

Combine new APIs into tlx.fence_async_shared() with a single tlx.fence(scope) API where scope
is a required argument: "gpu", "sys", or "async_shared".

Test plan:

  pytest python/test/unit/language/test_tlx.py::test_fence_gpu
  pytest python/test/unit/language/test_tlx.py::test_fence_sys
  pytest python/test/unit/language/test_tlx.py::test_descriptor_load
  pytest python/test/unit/language/test_tlx.py::test_descriptor_load_l2_cache_hint
  pytest python/test/unit/language/test_tlx.py::test_descriptor_store_l2_cache_hint
  pytest python/test/unit/language/test_tlx.py::test_descriptor_store_reduce
  pytest python/test/unit/language/test_tlx.py::test_descriptor_load_multicast
  pytest python/test/unit/language/test_tlx.py::test_dummy_layout_function_inlining
  pytest third_party/tlx/tutorials/testing/test_correctness.py

Internal use case D94839329

[meta-codesync]

@dshi7 has imported this pull request. If you are a Meta employee, you can view this in D94837956.

[htyu]

Thanks for working on this! Can we give the APIs a more general name? Since we already have tlx.fence_async_shared, maybe having a generic name like fence that covers both global and shared ?

[htyu]

Looks good overall. Can you please update the README?

[htyu]

LGTM!