[FIX] Extend the tracer's out-of-bounds guard to Gluon atomics, casts and async copies

[mark14wu]

Summary

Stacked on #491. This closes the three Gluon gaps that #491 left in the tracer's out-of-bounds guard:

• Pointer tables under Gluon. [FIX] Refuse out-of-bounds accesses in the tracer instead of corrupting memory #491 exempts every access in a program after it casts an integer to a pointer. Under Gluon that exemption never fired, so a valid pointer-table gather that mixes a kernel-argument row with a non-argument row was refused. Before [FIX] Refuse out-of-bounds accesses in the tracer instead of corrupting memory #491 it ran.
• Gluon atomics. gl.atomic_* and AMD buffer_atomic_* were never checked, so a straddling atomic still wrote past the tensor.
• Gluon async copies. These were never checked either. A gfx1250 shared_to_global overrun corrupted memory, and the Ampere and AMD global-to-shared copies read past the tensor.

Changes:

• tilelens/core/frontend/gluon.py (dispatch only).
  • Gluon builds atomics and int->pointer casts on TileLens's own simulation Builder, not on Triton's interpreter builder, so nothing patched them. A new GLUON_NAMESPACES entry, keyed by that Builder class, dispatches create_atomic_rmw, create_atomic_cas, create_buffer_atomic_rmw and create_int_to_ptr. Each gets a Gluon-specific subclass (GluonAtomicRMW, GluonAtomicCas, GluonBufferAtomicRMW, GluonIntToPtr) and an argument-normalising adapter. The buffer atomic adapter hands clients the per-lane pointers.
  • The subclasses are deliberate. The Sanitizer, RaceDetector and Profiler match op types exactly, so they keep running these ops concretely, as on main. Dispatching the plain AtomicRMW/AtomicCas/IntToPtr types was measured to crash the Sanitizer and RaceDetector on Gluon spin-lock and pointer-cast kernels that pass on main.
  • It dispatches Ampere async_copy.async_load / async_copy_global_to_shared, which were not intercepted at all. The async-copy adapters now accept CDNA4's (dest, ptr) keyword spelling. The TMA descriptor adapter now applies only to the TMA modules, so the Ampere functions, which share the TMA functions' names, no longer get the descriptor adapter.
• tilelens/clients/tracer/tracer.py (policy).
  • register_op_callback resolves callbacks along the op type's MRO. The Gluon subclasses above and the existing GluonAsyncCopyLoad / GluonAsyncCopyStore / GluonBufferLoadToShared now reach the tracer's existing Load/Store/atomic/IntToPtr callbacks.
  • Pointer and mask lanes are broadcast together before checking and recording. Gluon adapters see a builtin's operands before it broadcasts them, so an async copy can pass a row or scalar mask for a full tile.

Visualizer-visible change: async copies are now recorded. A sampled program's async copy appears as a Load tile (shared_to_global as a Store tile) with the copy's offsets and mask. Before, these copies were invisible. For example, an Ampere async_load of a 4×8 tile with a K-tail mask cols[None, :] < 6 now records a Load with offsets and mask both of shape (4, 8), and get_visualization_data() renders it.

Still not checked, as the README says:

• AMD buffer_load/buffer_store. Their adapters hand clients the base pointer, and fixing that changes the symbolic clients' buffer-op behaviour, so it needs its own PR.
• TMA/TDM descriptor copies, which the simulation bounds by the descriptor shape.
• Accesses that never touch a tensor argument, and programs after an int->pointer cast.

Inside gl.warp_specialize partitions, the tracer's per-program state (sampling, the int->pointer exemption, the program id in messages) is not inherited by the partition threads. That predates this PR and also affects gl.load/gl.store.

Test Plan

New tests:

• tests/end_to_end/test_gluon.py:
  • Out-of-bounds gl.atomic_add, gl.atomic_cas and cdna3.buffer_atomic_add are refused, and the sentinels are untouched. In-bounds and masked atomics still run.
  • All five global-to-shared async copies (Ampere ×2, CDNA4 ×2, gfx1250) run, record a Load, and refuse an overrun read. An overrunning gfx1250 shared_to_global is refused with the sentinels intact.
  • An async copy with a broadcast K-tail mask records shape-consistent Loads and visualizes.
  • The Gluon pointer-table gather is allowed.
  • Non-regression for the Sanitizer: a Gluon spin-lock kernel and a pointer bitcast after an int->pointer cast still run cleanly.
• tests/unit/test_adapters.py: the builder atomic adapters, the buffer atomic adapter (per-lane pointers; the ir.value() no-mask placeholder becomes None), the CDNA4 keyword spelling, and the Ampere async copies not getting the TMA adapter.
• tests/unit/test_tracer.py: op subclasses get their core op's callbacks.

Runs (CPU, Triton 3.8):

• Against [FIX] Refuse out-of-bounds accesses in the tracer instead of corrupting memory #491's source, all 14 new tracer/adapter tests fail. The Sanitizer non-regression tests pass there too, as they should. Removing the lane pairing or the TMA-namespace restriction fails its regression test.
• Sanitizer, RaceDetector and Profiler output on the investigated Gluon atomic, int->pointer and async-copy kernels is identical to [FIX] Refuse out-of-bounds accesses in the tracer instead of corrupting memory #491's tree.
• pytest tests/ -n 8: 397 passed, 3 skipped.

[github-actions]

Performance Benchmark

Benchmark	main (min)	PR (min)	Change	Samples
gemm	0.056s	0.055s	-1.0%	20 / 20
gemm_oob	0.061s	0.062s	+0.6%	20 / 20
indirect_load	0.011s	0.011s	+0.5%	20 / 20
nested_loop	0.116s	0.115s	-0.8%	20 / 20
block_pointer_loop_advance	0.061s	0.061s	-0.3%	20 / 20
liger_jsd	0.079s	0.078s	-1.0%	20 / 20
flaggems_layernorm	0.199s	0.196s	-1.8%	20 / 20
swiglu	0.094s	0.092s	-1.4%	20 / 20
cross_entropy	0.523s	0.504s	-3.7%	20 / 20
fused_linear_jsd	0.118s	0.119s	+0.8%	20 / 20
Total	1.318s	1.293s	-1.9%	N/A

Iterations: 1 warmup + 20 measured
Samples are shown as main / PR; long pytest benchmarks may use fewer samples.