Optimize CPU masked_scatter for contiguous arrays

[AK-Khan02]

Summary

Closes #3669.

Adds a CPU fast path for masked_scatter when mask, src, and out are row-contiguous. The existing ContiguousIterator implementation remains the fallback for strided and broadcasted cases.

Why

The CPU implementation previously used the general iterator path for all cases. For contiguous arrays, direct pointer indexing avoids per-element iterator/stride bookkeeping while preserving the same source-consumption semantics.

Local benchmark

CPU-only local benchmark, float32:

• 4M elements, 1% mask density: 5.746 ms -> 2.548 ms
• 4M elements, 10% mask density: 8.320 ms -> 4.467 ms
• 4M elements, 50% mask density: 23.816 ms -> 14.140 ms

Validation

PYTHONPATH=python/tests /tmp/mlx-mask-venv/bin/python -m unittest \
  python.tests.test_ops.TestOps.test_masked_scatter \
  python.tests.test_vmap.TestVmap.test_vmap_masked_scatter

[AK-Khan02]

WIP: this currently adds a scoped contiguous CPU fast path for masked_scatter with measurable local speedups. I’m still exploring whether a chunked prefix-count approach or other approaches could safely optimize the larger/general path further, so feedback on the current fast-path shape is welcome.