Allocation-free matrix construction + deprecate index helpers (#70)

[ctreffs]

Stack 4+5/5 — base: master. Also contains #70 (deprecate the public matrix index helpers), which was merged into this branch; see #70 for its description and tests.

Problem

• Matrix{2,3,4}.init(_ column0, …) built an Array of columns and handed it to Storage(columns), so every matrix construction allocated. That includes every non-simd multiply, which ends with Mat4x4d(Vec4d(…), …).
• init(_ values:) sliced the array into more arrays.
• Iteration allocated too: the non-simd storages used [column0, …].flatMap { $0 }, the simd quaternions built [x, y, z, w], and elements went through the type-erased AnyIterator.

Change

• New Storage*Protocol requirement: init(_ column0:, _ column1:, …). The simd matrix types (simd_float4x4 etc.) already have it, so their conformance needs no new code. The non-simd storages implement it, and their [Column] initializers delegate to it.
• Matrix column, array and values initializers all go through the column initializer, with no intermediate arrays. The [Vector] and [Value] initializers keep their preconditions.
• The non-simd storages drop their custom makeIterator. They are RandomAccessCollections and now use the default IndexingIterator<Self>.
• simd quaternions iterate their vector with SIMDScalarIterator.
• Matrix.elements is Array(storage), and SIMD.elements maps indices.

API note: the associated Iterator type of the non-simd storages and the simd quaternion storage changes, from IndexingIterator<[T]> to IndexingIterator<Self> / SIMDScalarIterator<SIMD4<T>>. This only breaks code that names those types explicitly.

Tests

MatrixConstructionTests checks, for all six matrix types, that the column, [column], [value] and array-literal initializers produce the same matrix in column-major layout, and that elements, Array(m) and for … in yield the same order. It also checks quaternion and SIMD element order. Both check-inlinable configurations still pass.

Benchmark

External consumer package, -c release, 5M mat4 * mat4:

	before (#68)	this PR
non-simd (Linux/Windows path)	156 ms	22 ms (≈7×)
simd	25.3 ms (median of 6)	25.3 ms (median of 6)

https://claude.ai/code/session_011WSUtoLptjea3s3iC5wTNC