Write DArray to HDF5/JLD without first converting to Array

It would be useful to have a way to write `DArray`s to file without first collecting all of the data on the local process. I found a pretty hacky way of doing this that looks like this:

``` julia
function write_darray{T<:AbstractFloat}(filepath::AbstractString, darr::DArray{T})
function write_localpart(pid::Int)
jldopen(filepath, "r+") do file
write(file, "inds$pid", collect(localindexes(darr)))
write(file, "arr$pid", localpart(darr))
end
end

# Write DArray metadata
jldopen(filepath, "w") do file
write(file, "dims", darr.dims)
write(file, "pids", collect(darr.pids))
end

# For each process, write the local indices and local part
for pid in darr.pids
remotecall_wait(pid, write_localpart, pid)
sleep(0.001) # This is for some reason necessary, or else HDF5 complains that the new object already exists
end
end

function read_darray(filepath::AbstractString)
file = jldopen(filepath, "r")
dims = read(file, "dims")
pids = read(file, "pids")
out = zeros(dims...)

# Reassemble local parts
for pid in pids
inds = read(file, "inds$pid")
out[inds...] = read(file, "arr$pid")
end

close(file)
return out
end
```

However, there must be a better way than this, especially one that would create a single variable in the file `filepath` (rather than `2n+2` in my case, where `n` is the number of processes the `DArray` is stored on).

Start with the issue's write_darray and read_darray examples, including jldopen, localindexes, localpart, and remotecall_wait. Investigate how DistributedArrays and the HDF5/JLD layer could represent one distributed variable rather than per-process objects. Done means a supported approach can write and read a DArray without collecting it locally and without creating 2n+2 file variables.

由索引模型根据 Issue 内容生成。