QR performance regression

[rayegun]

See JuliaLang/julia#61452 for history. This reduces to the following:

# Minimal reproducer: SubArray of SparseMatrixCSC loses specialized triangular solve
#
# `ldiv!(UpperTriangular(R), x)` is fast when R is a SparseMatrixCSC,
# but falls back to a generic O(n^2) method when R is a SubArray view.
#
# Introduced in JuliaSparse/SparseArrays.jl#676 (commit 95b6ac4).
# The new ldiv! uses a @view of F.R instead of indexing:
#   New (slow): https://github.com/JuliaSparse/SparseArrays.jl/blob/95b6ac4b0fa4d99e17920505dd36d7b95d91a0ab/src/solvers/spqr.jl#L513
#   Old (fast): https://github.com/JuliaSparse/SparseArrays.jl/blob/4500d8656d26a9328f50a48eaf9eca2f1fabc8ef/src/solvers/spqr.jl#L440

using SparseArrays, LinearAlgebra

n = 9000
R = spdiagm(0 => fill(2.0, n), 1 => fill(1.0, n - 1))
x = ones(n)

R_copy = R[Base.OneTo(n), Base.OneTo(n)]   # SparseMatrixCSC
R_view = @view R[Base.OneTo(n), Base.OneTo(n)]  # SubArray

# Warmup
ldiv!(UpperTriangular(R_copy), copy(x))
ldiv!(UpperTriangular(R_view), copy(x))

t_copy = @elapsed for _ in 1:10; ldiv!(UpperTriangular(R_copy), copy(x)); end
t_view = @elapsed for _ in 1:10; ldiv!(UpperTriangular(R_view), copy(x)); end

println("SparseMatrixCSC:  $(round(t_copy/10*1e6, digits=1)) μs")
println("SubArray view:    $(round(t_view/10*1e6, digits=1)) μs")
println("Slowdown:         $(round(t_view/t_copy, digits=0))x")

[JamesWrigley]

Eeek sorry about that, not sure how I missed it 😬 I'll have a go at fixing it on the weekend. From memory I think that line where F.R is indexed is the second biggest allocation in ldiv!() so it would be a bit unfortunate to go back to it, but maybe it's possible to allocate some other scratch buffers to copy that slice of F.R into to make another sparse matrix.