Add support for multiple layers of block arrays

It's sometimes useful to have multiple layers of blocks, see conversation in

https://github.com/JuliaDiffEq/MultiScaleArrays.jl/issues/25

I think one way to do this would be to support some sort of BlockView that converts an AbstractBlockVector{T} into an AbstractVector{Vector{T}}:

struct BlockView{AT, p, BA <: AbstractBlockArray} <: AbstractArray{AT, p}
   data::BA
end

_BlockView(data::BA) where BA<:AbstractBlockArray{T,p} where T = BlockView{Array{T,p},p,BA}(data)

getindex(A::BlockView, i...) = A[Block(i...)]
size(A::BlockView) = nblocks(A)

Then one create a hierarchical block array view (here we have sub-blocks of length 10 and super-blocks of length 5):

A_sub = PseudoBlockArray(randn(10*5*n), fill(10,5n))
A = PseudoBlockArray(BlockView(A_sub), fill(5,n))
A[Block(2)][3]  # returns the 3rd sub-block of the 2nd super-block

I'm still working out how to get it to work with A[Block(2)][Block(3)]...

Read the linked MultiScaleArrays discussion, then inspect AbstractBlockArray, PseudoBlockArray, BlockView, Block, and the existing indexing behavior. Define how hierarchical block arrays should be represented and how both A[Block(2)][3] and A[Block(2)][Block(3)] should behave; done means multiple block layers work consistently with the stated examples.