WIP: Generalize indexing fallbacks

[mbauman]

This is my take on #12486. Easiest to just show what this enables:

julia> immutable SymInt{T} <: AbstractArray{T,2}
          a::Vector{T}
          b::Vector{T}
          c::Vector{T}
       end
julia> Base.size(A::SymInt) = (length(A.a), 3)
julia> Base.getindex(A::SymInt, f::Symbol, i::Int) = getfield(A, f)[i]
julia> Base.checkbounds(A::SymInt, I::AbstractVector{Symbol}, J::AbstractVector{Int}) = true # punt

julia> A = SymInt([1,2,3,4],[11,12,13,14],[21,22,23,24]);

julia> A[[:a, :b, :c], [1,2,3]]
3x3 Array{Int64,2}:
  1   2   3
 11  12  13
 21  22  23

julia> using Interpolations

julia> itp = interpolate(reshape(1:20., 4, 5), BSpline{Linear}, OnGrid);

julia> itp[linspace(1,4,1001), linspace(1,5,1001)]
1001x1001 Array{Float64,2}: …

But there are two problems:

julia> A[:b, :]
ERROR: indexing SymInt{Int64} with types Tuple{Symbol,Colon} is not supported
 in error at ./error.jl:21
 in _getindex at abstractarray.jl:480
 in getindex at abstractarray.jl:462
 in eval at REPL.jl:3

julia> (1:100)[1.:100.]
WARNING: indexing with non Integer Reals is deprecated
 in depwarn at ./deprecated.jl:63
 in _unsafe_getindex at multidimensional.jl:205
 in getindex at abstractarray.jl:462
while loading no file, in expression starting on line 0
WARNING: indexing with non Integer Reals is deprecated
 in depwarn at ./deprecated.jl:63
 in _unsafe_getindex at multidimensional.jl:205
 in getindex at abstractarray.jl:462
while loading no file, in expression starting on line 0
WARNING: indexing with non Integer Reals is deprecated
 in depwarn at ./deprecated.jl:63
 in _unsafe_getindex at multidimensional.jl:205
 in getindex at abstractarray.jl:462
while loading no file, in expression starting on line 0
# … (100x)

The first commit is definitely correct. We should do that no matter what. The second commit means that we get N deprecation warnings for indexing with an N-element array of floats. And I don't know how to structure dispatch so we can pass along unknown non-Array types to multidimensional indexing without ending up with stack overflows.

I need to stop working on this for now, but I wanted to push this so y'all can see what I'm thinking.

[StefanKarpinski]

Would it make sense to make a separate PR for the commit that should definitely be done and just merge that as soon as tests pass?

[timholy]

I'll play around with this a bit more, but like you maybe later 😄.

[mbauman]

Superseded by #12540 and #12567.