Add cheap lu_instance/qr_instance for FixedSizeArrays

[devmotion]

Fixes #510.

lu_instance and qr_instance had cheap methods only for Matrix. A FixedSizeMatrix went through the generic fallback and ran a full factorization. LinearSolve's init calls these, so building a cache was about 50× slower than for a Matrix.

Changes

• Adds the extension FixedSizeArraysExt, which defines lu_instance and qr_instance for FixedSizeMatrix. Both factorize an empty similar(A, 0, 0) instead of A.
  • The real factorization types vary by eltype. For example, BigFloat gets Vector pivots and a QR instead of a QRCompactWY. Factorizing an empty matrix gives exactly the type of the real factorization, so these cases need no special handling.
  • qr_instance passes pivot on to qr. The generic fallback ignores it.
• Adds FixedSizeArrays as a weak dependency and test dependency with compat "1".
• Adds test/fixedsizearrays.jl. It checks Float64, Float32, ComplexF64, Int and BigFloat, with and without ColumnNorm(). Each test checks that the instance has the same type as the real factorization and that its factors are 0×0.

Benchmarks

64×64 FixedSizeMatrix{Float64}, Julia 1.13.1, FixedSizeArrays 1.3.0, Chairmarks @b:

	before (full factorization)	after
lu_instance(F)	19.3 μs, 3 allocs, 32.6 KiB	16 ns, 0 allocs
qr_instance(F)	50.5 μs, 6 allocs, 72.1 KiB	18 ns, 0 allocs

The Core test group passes locally, including Aqua.

🤖 Generated with Claude Code