[BlockSparseArrays] Redesign block views again

[mtfishman]

This PR is an iteration on the redesign of block views that was implemented in #1481, which implemented block views (i.e. @view a[Block(1, 1)]) as a SubArray wrapping the block sparse array and storing the location of the desired block. It is also related to #1498, which introduced view!/@view! that may instantiate blocks if they don't exist.

This PR defines a devoted type BlockView, and now @view a[Block(1, 1)] either directly outputs the data of the block if it exists, or a BlockView wrapping the block sparse array and the location of the desired block if the block doesn't exist.

So for example:

julia> using BlockArrays: Block

julia> using NDTensors.BlockSparseArrays: BlockSparseArray

julia> a = BlockSparseArray{Float64}([2, 3], [2, 3])
typeof(axes) = Tuple{BlockArrays.BlockedOneTo{Int64, Vector{Int64}}, BlockArrays.BlockedOneTo{Int64, Vector{Int64}}}

Warning: To temporarily circumvent a bug in printing BlockSparseArrays with mixtures of dual and non-dual axes, the types of the dual axes printed below might not be accurate. The types printed above this message are the correct ones.

2×2-blocked 5×5 BlockSparseArray{Float64, 2, Matrix{Float64}, NDTensors.SparseArrayDOKs.SparseArrayDOK{Matrix{Float64}, 2, NDTensors.BlockSparseArrays.BlockZero{Tuple{BlockArrays.BlockedOneTo{Int64, Vector{Int64}}, BlockArrays.BlockedOneTo{Int64, Vector{Int64}}}}}, Tuple{BlockArrays.BlockedOneTo{Int64, Vector{Int64}}, BlockArrays.BlockedOneTo{Int64, Vector{Int64}}}}:
 0.0  0.0  │  0.0  0.0  0.0
 0.0  0.0  │  0.0  0.0  0.0
 ──────────┼───────────────
 0.0  0.0  │  0.0  0.0  0.0
 0.0  0.0  │  0.0  0.0  0.0
 0.0  0.0  │  0.0  0.0  0.0

julia> a[Block(1, 1)] = randn(2, 2)
2×2 Matrix{Float64}:
  0.114978   3.56484
 -0.508884  -0.323752

julia> @view a[Block(1, 1)]
2×2 Matrix{Float64}:
  0.114978   3.56484
 -0.508884  -0.323752

julia> @view a[Block(2, 2)]
3×3 NDTensors.BlockSparseArrays.BlockView{Float64, 2, BlockSparseArray{Float64, 2, Matrix{Float64}, NDTensors.SparseArrayDOKs.SparseArrayDOK{Matrix{Float64}, 2, NDTensors.BlockSparseArrays.BlockZero{Tuple{BlockArrays.BlockedOneTo{Int64, Vector{Int64}}, BlockArrays.BlockedOneTo{Int64, Vector{Int64}}}}}, Tuple{BlockArrays.BlockedOneTo{Int64, Vector{Int64}}, BlockArrays.BlockedOneTo{Int64, Vector{Int64}}}}}:
 0.0  0.0  0.0
 0.0  0.0  0.0
 0.0  0.0  0.0

I think this will make some operations with block views easier to implement, and also help to simplify some code internally.

I'm not totally settled on the design that views of existing blocks directly returns the block data, an alternative would be that it also outputs a BlockView which is more comparable to the current behavior that it outputs a SubArray no matter what. However, I think it is worth trying out this design and seeing how it goes, it would be easy to change that later.

@ogauthe this is related to our discussion in #1336. I think this new design will make it a bit easier to fix some issues with block views you brought up, like using them in TensorAlgebra.contract and in Strided.@strided.

[mtfishman]

@ogauthe with this PR, this now works:

using BlockArrays: Block
using NDTensors.TensorAlgebra: contract
using NDTensors.BlockSparseArrays: BlockSparseArray
a = BlockSparseArray{Float64}([1], [1])
a[1, 1] = 1.0
b = @view a[Block(1, 1)]
contract(b, (1, 2), ones((1, 1)), (2, 3))

whether or not the block is instantiated, though my guess is that if the block isn't instantiated it may not be doing something efficient.