CI: Try removing the NumPy <2.5 pin from the numba-cuda-mlir nightly

[rwgk]

Description

The nightly-numba-cuda-mlir install path currently adds numpy<2.5 in ci/tools/run-tests. The ci-nightly.yml jobs select that install path through the nightly-numba-cuda-mlir test mode.

The constraint was introduced in #2283 as a workaround for NVIDIA/numba-cuda-mlir#154: numba-cuda-mlir 0.3.0 and 0.4.0 tried to register np.row_stack, which NumPy 2.5 removed. NVIDIA/numba-cuda-mlir#146 fixed the version guard, and the fix shipped in numba-cuda-mlir v0.4.1.

The nightly install does not pin numba-cuda-mlir to 0.4.0; it installs a compatible release and then checks out the matching release tag to run its tests. We should try removing the NumPy cap so this nightly lane exercises current NumPy releases again.

xref: this was discovered in while working on nvbug 6665297

Proposed change

• Remove only the "numpy<2.5" argument and its associated workaround comment from the nightly-numba-cuda-mlir branch in ci/tools/run-tests.
• Leave the separate nightly-numba-cuda path and all other dependency constraints, skips, and test commands unchanged. This change does not assert NumPy 2.5 compatibility for the legacy numba-cuda package.

Validation

• Run all currently configured nightly-numba-cuda-mlir jobs: Python 3.12 on Linux and Windows with CUDA 12.9.1 and 13.3.0.
• Confirm the existing package listing shows NumPy 2.5 or newer and numba-cuda-mlir 0.4.1 or newer.
• Compare with the current nightly baseline and confirm there is no np.row_stack error or other new NumPy-related failure; this experiment need not resolve unrelated existing failures.
• If a job exposes a real remaining incompatibility, record the resolved package versions and failure, then add back only the narrowest justified constraint with an upstream reference.

[sylvesterkaczmarek]

Implementation review

I reviewed #2803 against the current main implementation and the original NumPy workaround.

The numpy<2.5 constraint is now stale. It was introduced for numba-cuda-mlir 0.4.0 because of its np.row_stack usage. That compatibility issue was fixed in 0.4.1, while the nightly path resolves a current compatible release.

The proposed change remains strictly scoped to:

• removing numpy<2.5 from nightly-numba-cuda-mlir
• removing the associated obsolete workaround comment
• leaving the legacy nightly-numba-cuda path and all other install arguments unchanged

Validation completed on Linux and Windows. Twenty-four original-versus-patched install-path comparisons per platform passed, along with pip check and NumPy 2.5.3 / numba-cuda-mlir 0.5.2 compatibility checks.

The remaining validation is the repository's four GPU nightly jobs for CUDA 12.9.1 and 13.3.0 on Linux and Windows.

Review conclusion: the scoped change is ready for GPU CI validation. A clean implementation branch is prepared. Please assign the issue if you are happy for me to open the PR.