Implement vector indexing for SDiagonal with statically sized indices

[jishnub]

On master:

julia> m = SDiagonal(@SVector [11, 12, 13, 14])
4×4 Diagonal{Int64, SVector{4, Int64}}:
 11   ⋅   ⋅   ⋅
  ⋅  12   ⋅   ⋅
  ⋅   ⋅  13   ⋅
  ⋅   ⋅   ⋅  14

julia> axes(m)
(Base.OneTo(4), Base.OneTo(4))

julia> @btime $m[:, 1]
  256.280 ns (4 allocations: 256 bytes)
4-element SparseArrays.SparseVector{Int64, Int64} with 1 stored entry:
  [1]  =  11

julia> @btime $m[:,:]
  714.179 ns (5 allocations: 384 bytes)
4×4 SparseArrays.SparseMatrixCSC{Int64, Int64} with 4 stored entries:
 11   ⋅   ⋅   ⋅
  ⋅  12   ⋅   ⋅
  ⋅   ⋅  13   ⋅
  ⋅   ⋅   ⋅  14

julia> @btime $m[:]
  521.792 ns (4 allocations: 256 bytes)
16-element SparseArrays.SparseVector{Int64, Int64} with 4 stored entries:
  [1 ]  =  11
  [6 ]  =  12
  [11]  =  13
  [16]  =  14

After this PR:

julia> axes(m)
(SOneTo(4), SOneTo(4))

julia> @btime $m[:,1]
  2.676 ns (0 allocations: 0 bytes)
4-element SVector{4, Int64} with indices SOneTo(4):
 11
  0
  0
  0

julia> @btime $m[:,:]
  5.316 ns (0 allocations: 0 bytes)
4×4 SMatrix{4, 4, Int64, 16} with indices SOneTo(4)×SOneTo(4):
 11   0   0   0
  0  12   0   0
  0   0  13   0
  0   0   0  14

julia> @btime $m[:]
  18.406 ns (0 allocations: 0 bytes)
16-element SVector{16, Int64} with indices SOneTo(16):
 11
  0
  0
  0
  0
 12
  0
  0
  0
  0
 13
  0
  0
  0
  0
 14

[jishnub]

Speialized copy(::SDiagonal) in d1eb08e.

Before:

julia> m = SDiagonal(@SVector [11, 12, 13, 14])
4×4 Diagonal{Int64, SVector{4, Int64}} with indices SOneTo(4)×SOneTo(4):
 11   ⋅   ⋅   ⋅
  ⋅  12   ⋅   ⋅
  ⋅   ⋅  13   ⋅
  ⋅   ⋅   ⋅  14

julia> @btime [copy($m) for i = 1:100];
  1.155 μs (101 allocations: 5.56 KiB)

After:

julia> @btime [copy($m) for i = 1:100];
  809.139 ns (1 allocation: 3.25 KiB)

[jishnub]

Test failure on nightly seems unrelated (ambiguity in lu). Gentle bump

[mateuszbaran]

This is the same as the generic fallback:

StaticArrays.jl/src/abstractarray.jl

Line 190 in 29a76ec

@inline reshape(a::AbstractArray, s::Tuple{SOneTo,Vararg{SOneTo}}) = reshape(a, homogenize_shape(s))

.
Do we really need this method?

[mateuszbaran]

Other than this small point, the PR looks nice 👍 .