Speed up in-place vector-to-C-contiguous-matrix broadcast on CPU

[vchamarthi]

Speed up in-place vector-to-C-contiguous-matrix broadcast on CPU

In-place binary elementwise ops that broadcast a vector against a C-contiguous matrix
(m += row, m += col[:, None]) fell through to the general strided kernel (scalar, one
element per work-item) on the CPU device, even though a vectorized broadcast kernel already
exists and is used by the out-of-place path.

Changes

• Row broadcast (m += row): add the missing C-contiguous dispatch branch to
  py_binary_inplace_ufunc (the out-of-place path already had it; the in-place template only
  had the F-style {1,0} branch). Reuses the existing BinaryInplaceRowMatrixBroadcastingFunctor,
  so it benefits all in-place binary ufuncs — no new kernel.
• Column broadcast (m += col[:, None]): add BinaryInplaceColMatrixBroadcastingFunctor
  (mat[gid] += vec[gid / n1]) and wire it for add via a defaulted extra template
  parameter, leaving all other in-place ufuncs unchanged.

Both paths are guarded by exact simplified-stride checks and fall back to the strided kernel
otherwise; results are bitwise-identical.

Results

13th Gen Intel Core i5-13400 (CPU/OpenCL), float32, dpnp built from source. Same op, before/after
this change:

op (D = 21846x21846)	before	after
D += row	247 ms	58 ms
D += col[:, None]	252 ms	150 ms

This was the only op category where dpnp-on-CPU lost to NumPy in dpbench pairwise_distance
(M16Gb, single). End-to-end via the dpbench CLI (--repeat 30, --validate passing):
0.92x → 1.52x vs stock NumPy (731 ms → 445 ms; NumPy 678 ms).

Checklist

• Have you provided a meaningful PR description?
• Have you added a test, reproducer or referred to an issue with a reproducer?
  (TestAdd::test_inplace_row_broadcast, TestAdd::test_inplace_column_broadcast)
• Have you tested your changes locally for CPU and GPU devices? (row + column broadcast verified
  bitwise-equal to NumPy on opencl:cpu and level_zero:gpu, across float32/64 + int32/64 and
  shapes incl. non-sub-group-multiple row lengths)
• Have you made sure that new changes do not introduce compiler warnings? (clean recompile with
  icpx 2026.0, -Wall -Wextra: 0 warnings)
• Have you checked performance impact of proposed changes?
• Have you added documentation for your changes, if necessary? (no public API change)
• Have you added your changes to the changelog?

[coveralls]

coverage: 78.073%. remained the same — vchamarthi:fix/inplace-cmatrix-broadcast-cpu into IntelPython:master