[samridh10exe]

Description

Array.slice() failed type checking in cuda.lang for any start and stop, including integer literals. The shared implementation expects 0D tiles and builds the view with cuTile pointer ops, while cuda.lang uses scalars and its own pointers.

Add a cuda.lang implementation that builds the view with PointerOffset and Array.from_parts, like array_getitem, and keeps the same checks on start and stop. The new tests in test_slice.py cover constant bounds, runtime bounds on the last axis, and the rejected cases.

I couldn't run the GPU tests, since mlir2cubin is not in the public build. On CPU, the kernels from the new tests fail type checking before this change and lower to IR after it, with the view offset at start * stride and the axis size stop - start.

closes #101

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.