[COMPAT] Require Triton 3.8

[mark14wu]

Summary

Raises the minimum Triton version from 3.6.0 to 3.8.0 and removes the compatibility code that only existed for older releases.

The old triton>=3.6.0 floor was already inaccurate: on Triton 3.6, import tilelens.core.simulation.gluon fails because triton.experimental.gluon.language.amd.gfx1250 has no cluster there. CI has been resolving Triton 3.8.0 with torch 2.14.1 (torch 2.14 pins triton~=3.8.0), so 3.8 is what is actually tested.

Changes

• pyproject.toml: triton>=3.6.0 → triton>=3.8.0.
• Gluon simulator and frontend: replace the guarded imports of amd.gfx1250 (and its async_copy / cluster / mbarrier / tdm) and nvidia.blackwell.clc with plain imports, and remove the None handling that depended on them (get_partitioned_shared_layout, _GLUON_BUILTIN_MODULES, _GLUON_BUILTIN_CLASSES, _GLUON_NON_SEMANTIC_BUILTINS, _existing_ops).
• Gluon simulator: import _mxfp_value_handle_to_float32 and _unpack_e2m1 from triton.runtime.interpreter and delete the local fallback copies (Triton 3.6 did not have these helpers; 3.8 does).
• Triton frontend: snapshot tl.core.tensor_descriptor_base without a hasattr guard. Update the comments that cited <3.6 support. TileLens keeps its own snapshot scope because it covers both tl and tl.core for nested and generated kernels, while Triton's scope only records the modules visible from fn.
• Tests: remove skips that never fire on 3.8: CDNA4 async copy, TMA im2col (×3), fp4-padded tensor memory, and the tensor_descriptor_base patch-scope test.

On Triton 3.8.0, each removed fallback's primary path exists. The Ampere async_copy_global_to_local skip in test_gluon.py is unchanged: that name does not exist in 3.6 or 3.8, so it is not a version shim.

Testing

• Checked on Triton 3.8.0 (torch 2.14.1) that every removed import or attribute fallback resolves to its primary path, and that tilelens.core.frontend.gluon and tilelens.core.simulation.gluon import cleanly.
• pre-commit passes on the changed files (ruff, ruff-format, mypy, codespell).
• pytest --collect-only on tests/end_to_end/test_gluon.py and tests/unit/test_patch_scope.py works.
• I did not run the full test suite locally; this PR's CI covers it on Triton 3.8.0.

Follow-ups

• Open PR [FEAT] Add the host-compiled IR lifecycle for clients #479 currently targets a Triton 3.6 + 3.8 window; after this lands, it can drop 3.6.

[github-actions]

Performance Benchmark

Benchmark	main (min)	PR (min)	Change	Samples
gemm	0.105s	0.105s	+0.0%	20 / 20
gemm_oob	0.117s	0.116s	-0.2%	20 / 20
indirect_load	0.022s	0.022s	-0.9%	20 / 20
nested_loop	0.237s	0.235s	-0.9%	20 / 20
block_pointer_loop_advance	0.125s	0.126s	+1.0%	20 / 20
liger_jsd	0.139s	0.139s	+0.3%	20 / 20
flaggems_layernorm	0.396s	0.397s	+0.2%	20 / 20
swiglu	0.169s	0.171s	+0.9%	20 / 20
cross_entropy	0.977s	0.980s	+0.3%	20 / 20
fused_linear_jsd	0.211s	0.210s	-0.3%	20 / 20
Total	2.496s	2.500s	+0.2%	N/A

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