Various fixes for propagating CachedArrayStyle correctly

[DanielVandH]

Now going to use this PR to get various other fixes for propagating CachedArrayStyle. Makes it easier for me when debugging but if it's preferred I can later break it up into smaller PRs. Once all the changes are in, I'll edit in all the new features here. Original PR text below.

---

This changes Vcat and Hcat to use the combined BroadcastStyle between a LazyArrayStyle and the combined style of its arguments. This fixes e.g. Vcat of cached arguments not being recognised as a CachedArrayStyle.

The implementation needs to use some recursion to get the BroadcastStyle of the arguments via tuple_type_broadcastlayout, but basically it's just doing the same as what you'd get from result_style(LazyArrayStyle{N}(), combine_style(A.args...)) where A is a Vcat/Hcat

---

[codecov]

Codecov Report

❌ Patch coverage is 0% with 32 lines in your changes missing coverage. Please review.
✅ Project coverage is 0.00%. Comparing base (8bf603d) to head (6905338).
⚠️ Report is 1 commits behind head on master.

Files with missing lines	Patch %	Lines
src/lazyoperations.jl	0.00%	13 Missing ⚠️
src/lazyconcat.jl	0.00%	10 Missing ⚠️
ext/LazyArraysBandedMatricesExt.jl	0.00%	6 Missing ⚠️
src/lazyapplying.jl	0.00%	2 Missing ⚠️
src/lazybroadcasting.jl	0.00%	1 Missing ⚠️

❗ There is a different number of reports uploaded between BASE (8bf603d) and HEAD (6905338). Click for more details.

HEAD has 7 uploads less than BASE

Flag	BASE (8bf603d)	HEAD (6905338)
	13	6

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #391       +/-   ##
==========================================
- Coverage   95.90%   0.00%   -95.91%     
==========================================
  Files          17      17               
  Lines        3372    3348       -24     
==========================================
- Hits         3234       0     -3234     
- Misses        138    3348     +3210     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[DanielVandH]

Similar cases missing a CachedArray style also show up in e.g. BlockArrays and LazyBandedMatrices:

julia> v = Accumulate(*, 1:10);

julia> (Base.BroadcastStyle ∘ typeof).((v, unitblocks(v), BlockBroadcastArray(vcat, unitblocks(v), unitblocks(v))))
(LazyArrays.CachedArrayStyle{1}(), BlockArrays.BlockedStyle{1}(), LazyArrays.LazyArrayStyle{1}())

So I might have to eventually go and fix these, assuming it would be appropriate that these all return CachedArrayStyle{1}()? Similarly,

julia> Base.BroadcastStyle(typeof(vec(Vcat(v', v')))) # interlaced
Base.Broadcast.DefaultArrayStyle{1}()

will get a fix (this last one is now fixed in JuliaLinearAlgebra/ArrayLayouts.jl#268)

[dlfivefifty]

I guess we don't need a CachedBandedArrayStyle because CachedStyle lowers to the data after resizing?

[DanielVandH]

I think that's right. I wasn't too sure exactly about cases like that (and e.g. the block matrices like I mentioned in my previous comment), but I think I reasoned similarly that CachedArrayStyle is fine

[DanielVandH]

This QRCompactWYQ type always seems to run into this issue. I'm not sure how to get around it properly this time, since there would need to be somewhere to call broadcastable earlier (which converts it into a Matrix), since by the time the calls get into using type information only it's too late. If LinearAlgebra just had

BroadcastStyle(::Type{<:LinearAlgebra.QRCompactWYQ}) = DefaultArrayStyle{2}()
BroadcastStyle(::Type{<:LinearAlgebra.AdjointQ}) = DefaultArrayStyle{2}()

it would be fine, but can't do that on our side without piracy. I don't really understand why they don't already do this.

Solutions like defining an internal _BroadcastStyle or overloading combine_styles seem like code smell. Have you had to deal with something similar before @dlfivefifty? An MWE for this is

julia> using LazyArrays, LinearAlgebra

julia> B = LazyArrays.CachedArray(randn(3, 3)); Q = qr(randn(3, 3)).Q; collect(B); Q * B ≈ Q * B.data
ERROR: MethodError: no method matching ndims(::Type{LinearAlgebra.QRCompactWYQ{Float64, Matrix{Float64}, Matrix{Float64}}})
The function `ndims` exists, but no method is defined for this combination of argument types.

Closest candidates are:
  ndims(::Type{Union{}}, Any...)
   @ Base abstractarray.jl:276
  ndims(::Type{<:Ref})
   @ Base refpointer.jl:104
  ndims(::Type{<:Applied{<:Any, typeof(*), Args}}) where Args
   @ LazyArrays c:\Users\djv23\.julia\dev\LazyArrays.jl\src\linalg\mul.jl:41
  ...

Stacktrace:
  [1] Base.Broadcast.BroadcastStyle(::Type{LinearAlgebra.QRCompactWYQ{Float64, Matrix{Float64}, Matrix{Float64}}})
    @ Base.Broadcast .\broadcast.jl:103
  [2] tuple_type_broadcastlayout
    @ c:\Users\djv23\.julia\dev\LazyArrays.jl\src\lazyconcat.jl:445 [inlined]
  [3] applybroadcaststyle(::Type{ApplyArray{Float64, 2, typeof(*), Tuple{…}}}, ::ApplyLayout{typeof(*)})
    @ LazyArrays c:\Users\djv23\.julia\dev\LazyArrays.jl\src\lazyapplying.jl:334
  [4] Base.Broadcast.BroadcastStyle(M::Type{ApplyArray{Float64, 2, typeof(*), Tuple{…}}})
    @ LazyArrays c:\Users\djv23\.julia\dev\LazyArrays.jl\src\lazyapplying.jl:336
  [5] combine_styles(c::ApplyArray{Float64, 2, typeof(*), Tuple{LinearAlgebra.QRCompactWYQ{…}, CachedArray{…}}})
    @ Base.Broadcast .\broadcast.jl:433
  [6] combine_styles(c1::ApplyArray{Float64, 2, typeof(*), Tuple{…}}, c2::Matrix{Float64})
    @ Base.Broadcast .\broadcast.jl:438
  [7] broadcasted
    @ .\broadcast.jl:1353 [inlined]
  [8] broadcast_preserving_zero_d
    @ .\broadcast.jl:882 [inlined]
  [9] -(A::ApplyArray{Float64, 2, typeof(*), Tuple{LinearAlgebra.QRCompactWYQ{…}, CachedArray{…}}}, B::Matrix{Float64})
    @ Base .\arraymath.jl:8
 [10] isapprox(x::ApplyArray{…}, y::Matrix{…}; atol::Int64, rtol::Float64, nans::Bool, norm::typeof(norm))
    @ LinearAlgebra C:\Users\djv23\.julia\juliaup\julia-1.12.1+0.x64.w64.mingw32\share\julia\stdlib\v1.12\LinearAlgebra\src\generic.jl:1982
 [11] isapprox(x::ApplyArray{Float64, 2, typeof(*), Tuple{LinearAlgebra.QRCompactWYQ{…}, CachedArray{…}}}, y::Matrix{Float64})
    @ LinearAlgebra C:\Users\djv23\.julia\juliaup\julia-1.12.1+0.x64.w64.mingw32\share\julia\stdlib\v1.12\LinearAlgebra\src\generic.jl:1978
 [12] top-level scope
    @ REPL[100]:1
Some type information was truncated. Use `show(err)` to see complete types.

[dlfivefifty]

I think

defining an internal _BroadcastStyle

is probably the best way around this?

I have played around in the past with making a type so that Q's are <: AbstractMatrix (I think AbstractQ is kinda poorly thought out btw):

JuliaLinearAlgebra/MatrixFactorizations.jl#69