lazy module loading

[CarloLucibello]

Fix #126 using FileIO and a similar method for lazy loading modules.
Time-to-first-MNIST goes from 14.5s to 5.5s.

###### MASTER ##############

julia> @time using MLDatasets
 12.421854 seconds (19.82 M allocations: 1.144 GiB, 6.86% gc time, 61.53% compilation time)

julia> @time MNIST()
  1.965113 seconds (1.34 M allocations: 301.964 MiB, 13.25% gc time, 44.99% compilation time)
dataset MNIST:
  metadata    =>    Dict{String, Any} with 3 entries
  split       =>    :train
  features    =>    28×28×60000 Array{Float32, 3}
  targets     =>    60000-element Vector{Int64}

#### THIS PR #####################
julia> @time using MLDatasets
  3.757601 seconds (5.07 M allocations: 293.185 MiB, 5.89% gc time, 66.09% compilation time)

julia> @time MNIST()
  1.653665 seconds (1.33 M allocations: 299.937 MiB, 3.27% gc time, 48.02% compilation time)
dataset MNIST:
  metadata    =>    Dict{String, Any} with 3 entries
  split       =>    :train
  features    =>    28×28×60000 Array{Float32, 3}
  targets     =>    60000-element Vector{Int64}

[CarloLucibello]

@johnnychen94 I'm running in some world age issue, see the failing tests. Do you know how to fix them?

[codecov-commenter]

Codecov Report

Merging #128 (53a6ab7) into master (e3cc061) will decrease coverage by 3.61%.
The diff coverage is 43.47%.

@@            Coverage Diff             @@
##           master     #128      +/-   ##
==========================================
- Coverage   33.26%   29.64%   -3.62%     
==========================================
  Files          38       39       +1     
  Lines        1527     1565      +38     
==========================================
- Hits          508      464      -44     
- Misses       1019     1101      +82     

Impacted Files	Coverage Δ
src/MLDatasets.jl	100.00% <ø> (ø)
src/containers/tabledataset.jl	0.00% <0.00%> (-89.29%)	⬇️
src/datasets/graphs/reddit.jl	4.25% <0.00%> (ø)
src/datasets/text/udenglish.jl	56.86% <ø> (ø)
src/datasets/vision/cifar10.jl	2.40% <0.00%> (-0.03%)	⬇️
src/datasets/vision/cifar100.jl	2.40% <ø> (ø)
src/datasets/vision/emnist.jl	10.00% <0.00%> (ø)
src/datasets/vision/svhn2.jl	1.96% <0.00%> (ø)
src/abstract_datasets.jl	22.58% <37.50%> (-1.56%)	⬇️
src/require.jl	37.50% <37.50%> (ø)
... and 14 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update e3cc061...53a6ab7. Read the comment docs.

[johnnychen94]

The overall change looks good to me although I'm not confident that users want the @require error -- I understand how you want to reduce the package loading latency, just unsure if it's a good direction.

Said that, if we introduce the "hard" requirement on package loading, then we might need to add some explaination before/after the https://juliaml.github.io/MLDatasets.jl/stable/#Basic-Usage to tell users how to properly handle this.

Because the hard requirement will break people's code, this will need a 0.7.0 version bump.