Add support for reading H5T_ARRAY

[timholy]

This gets us partway to fixing #50. @simonster , if you have time to review, I would greatly value your opinion. The most "controversial" aspect of this is that I decided to abandon my previous attempts to read fixed-size arrays by encoding the dimensions within an "instance" and instead decided to create a stub-type for each fixed-size array (with the dimensions encoded as parameters, e.g., FixedArray3{Float64, 3, 5, 7} would correspond to Array{Float64,3} of size (3,5,7). The stub types don't actually allocate any memory.

The reason I went with the type-based approach is because the whole infrastructure around hdf5_to_julia_eltype etc is built on types rather than instances, and arrays-of-fixedarrays (as in the example file I added to readremote) seemed ugly if implemented with instances.

But see if you agree.

[simonster]

I'll look at this more closely later, but generally I think it's better if we can avoid dynamically generating types. One possibility is to use:

immutable Dim{N} end
immutable FixedArray{T,N<:(Dim...)} end

Then we have:

param{N}(::Type{Dim{N}}) = N
size{T,D}(::Type{FixedArray{T,D}}) = map(param, D)::(Int...)

size(FixedArray{Float64,(Dim{3},Dim{4},Dim{7})}) == (3,4,7)

There might even be a way to avoid the Dim type, but I don't see it ATM.

[timholy]

Very interesting. I tried the same thing without the Dim wrapper, and got this:

julia> immutable FixedArray2{T,N<:(Int...)} end

julia> F = FixedArray2{Float64,(3,4,2)}
ERROR: type: apply_type: in FixedArray2, expected Type{T<:Top}, got (Int64,Int64,Int64)

I should have thought to do the wrapper. Nice work!

I completely agree that it's better to avoid dynamically-generating types. I don't think you have to review this in detail, I'll generate a version based on your suggestion.

[timholy]

CC @JeffBezanson in case he's interested in the tuple-of-types vs tuple-of-integers issue.