[language](fix) Allow non-power-of-two arange ranges

[Ritsuka314]

Rationale

AscendNPU-IR now supports non-power-of-two tensor shapes, but the Triton Ascend patch still rejects non-power-of-two tl.arange ranges in SIMT mode before they reach lowering. The same patch already relaxes the generic Python block-shape and C++ tensor-size restrictions.

Remove the remaining semantic guard so both SIMD and SIMT compilation can use the downstream non-power-of-two support.

This is the Triton-Ascend frontend companion to AscendNPU-IR !1357, which adds the corresponding downstream support and regression coverage.

Validation

• git apply --check third_party/ascend/patch/triton-ascend-3.6.0.patch
• pre-commit run --from-ref github-upstream/main-dev --to-ref HEAD
• Build and device execution were not run; this machine has no Ascend NPU, and submodule initialization was intentionally skipped.

New contributor declaration

• I am not making a trivial change, such as fixing a typo in a comment.
• I have written a PR description following these rules: https://cbea.ms/git-commit/#why-not-how.
• I have run pre-commit from the PR base to HEAD.
• This PR does not add a test because it removes a frontend rejection whose downstream behavior is covered by the linked AscendNPU-IR feature validation.
• I have not added any lit tests.

[github-actions]

✅ OpenCodeReview: Review skipped: no items were selected.