Fix the CPU scan over a size one axis with a padded stride

[kapellirohith]

Proposed changes

A scan over a size one axis that carries a padded stride leaves the output
unwritten, so it returns whatever was in the freshly allocated buffer:

x = mx.arange(1, 11, dtype=mx.float32).reshape(1, 10)[:, 2:]
mx.cumsum(x, axis=0)

The strided CPU scan takes its row count from
size / shape[axis] / strides[axis]. row_contiguous ignores size one axes,
so such an axis can carry any stride and the array is still flagged row
contiguous. Here the shape is (1, 8) with strides (10, 1), the count
floors to zero, and the loop body never runs.

A padded stride is only possible on a size one axis, and a scan over a size
one axis is elementwise, which is what the contiguous path with a stride of
one already computes, so no copy is needed.

Introduced by c423074 "redesign for faster cpu/gpu synch (#1869)"
(2025-03-06).

The GPU backends are not affected. Both Metal and CUDA give the output the
input's strides and flags, so the two stay in the same frame, and their row
count comes from data_size() rather than the stride.

Note for whoever lands second: #3907 already fails git apply --check against
pristine main at scan.cpp:200, independently of this change. The resolution
keeps both its AccT init and the || in.shape(axis) == 1 here. The two are
semantically orthogonal.

Ran locally on an M3 Pro: the Python suite on the GPU and with
mx.set_default_device(mx.cpu), the C++ suite on both devices, pre-commit,
and the new test five times on each device. Not covered locally: CUDA, Linux,
Windows.