Guard scatter_axis against 64-bit outputs on the GPU

[obchain]

Proposed changes

Fixes #3690.

mx.put_along_axis / scatter_add_axis with a 64-bit element dtype (int64/uint64) fail the Metal library JIT build instead of raising a clean "unsupported dtype" error. In mlx/backend/metal/kernels/atomic.h, packing_size<T> = sizeof(uint)/sizeof(T) is 0 for 8-byte T, so uint_or_packed<T> declares a zero-length array (hard C++ error) and offset / packing_size<T> divides by zero. The whole mlx-metallib build then fails.

The plain Scatter path already guards 8-byte outputs on the GPU — in scatter() (ops.cpp) and Scatter::eval_gpu (indexing.cpp) — but ScatterAxis was missing the equivalent guard. This adds it in both places, mirroring Scatter:

• scatter_axis() in mlx/ops.cpp — raises on GPU for 8-byte dtypes, matching the existing scatter() guard. CPU is unaffected.
• ScatterAxis::eval_gpu in mlx/backend/metal/indexing.cpp — same guard as Scatter::eval_gpu.

Repro before the fix (Metal device):

import mlx.core as mx
mx.set_default_device(mx.gpu)
x   = mx.zeros((4, 8), dtype=mx.int64)
idx = mx.array([[0],[1],[2],[3]])
upd = mx.ones((4, 1), dtype=mx.int64)
mx.eval(mx.put_along_axis(x, idx, upd, axis=1))  # was: Metal JIT build failure

After the fix this raises a clean ValueError on the GPU, and continues to work on the CPU.

Checklist

Put an x in the boxes that apply.

• I have read the CONTRIBUTING document
• I have run pre-commit run --all-files to format my code / installed pre-commit prior to committing changes
• I have added tests that prove my fix is effective or that my feature works
• I have updated the necessary documentation (if needed)

[zcbenz]

The check in metal/indexing.cpp along is enough, and it actually works in the cuda backend, the test should also be updated for metal only