[tivris]

Follow-up to #4744. concatenate_gpu built destination views with copy_shared_buffer(..., out_slice.size(), data_offset), so data_size was the input element count rather than its span in the output. When concatenating along axis 1 or later into an output past 2^31 elements (mx.stack along axis 1 or later uses the same path), the copy picks 32-bit indices and wraps: Metal silently leaves stale memory where late rows should be, and CUDA is expected to fault.

In both Metal and CUDA backends, views are now built using slice_gpu, which computes the true span. The CPU backend is already correct and untouched.

test_concatenate_large_offsets concatenates along axis 1 to produce a (2^27 + 1, 16) uint8 array and checks the last row. It fails on main on GPU and passes with the fix on GPU and CPU. Neighboring ops tests and pre-commit pass. CUDA was compile-checked only, with warnings as errors on CUDA 12.9 and 13.0.

• ☑️ I understand it is strictly prohibited to use AI to write PR description
• AI usage disclosure: debugging, code and tests