NDArray M4: Reduction Operations

[Copilot]

Implements Milestone 4 — reduction operations for NDArray[Double]: full reductions, axis reductions, arg reductions, and linear algebra (dot/matmul), as specified in the design doc.

New: vecxt/src/ndarrayReductions.scala

Full reductions — return Double, col-major fast path delegates to existing SIMD/BLAS Array[Double] extensions:

• sum, mean, min, max, product, variance, norm
• argmax, argmin (flat col-major index)

Axis reductions — output shape removes the reduced dimension (NumPy keepdims=False semantics):

• sum(axis), mean(axis), min(axis), max(axis), product(axis), argmax(axis), argmin(axis)

Linear algebra:

• dot — 1-D only, delegates to BLAS ddot on col-major inputs
• matmul / @@ — 2-D × 2-D, materialises non-contiguous inputs then delegates to BLAS dgemm via Matrix

val a = NDArray(Array(1.0, 2.0, 3.0, 4.0, 5.0, 6.0), Array(2, 3))

a.sum          // 21.0
a.mean         // 3.5
a.argmax       // 5
a.sum(axis = 0)  // NDArray shape [3]: [3.0, 7.0, 11.0]
a.sum(axis = 1)  // NDArray shape [2]: [9.0, 12.0]

val b = NDArray(Array(7.0, 8.0, 9.0, 10.0, 11.0, 12.0), Array(3, 2))
a @@ b  // NDArray shape [2,2]: [[76, 103], [100, 136]]

Implementation notes

• General strided kernel iterates in column-major coordinate order — same pattern as NDArrayDoubleOps.
• All SIMD acceleration is inherited by delegating col-major full reductions to the existing arrays.* extensions; no platform-specific code required.
• NDArrayReductionHelpers is a private top-level object (not nested inside NDArrayReductions) to work around a Scala 3 resolution issue where extension methods defined in an object shadow same-named extensions from imports, even for different receiver types.

Other changes

• vecxt/src/all.scala — exports NDArrayReductions.*
• vecxt/test/src/helpers.forTests.scala — adds assertNDArrayShapeAndClose
• vecxt/test/src/ndarrayReductions.test.scala — cross-platform tests covering all operations, 2-D and 3-D axis reductions (with hand-verified values from the design doc), strided/transposed paths, and error cases