Standardize missing value representation/handling logic

[takuti]

Close #20

• Missing values are represented by a custom Unknown type, which consists of Missing and Nothing.
• Both nothing and missing values in input data are replaced with 0 at the DataAccessor level to avoid the complication in the later processes.
• Each recommender assumes the data is always a float (sparse) matrix and filled by actual numbers. 0.0 is specially treated as needed.
• Event.value doesn't accept the unknown values; only Infinite types are considered.

[codecov-commenter]

Codecov Report

Merging #24 (2d590fe) into master (551019e) will increase coverage by 0.04%. The diff coverage is 98.83%.

@@            Coverage Diff             @@
##           master      #24      +/-   ##
==========================================
+ Coverage   97.23%   97.28%   +0.04%     
==========================================
  Files          19       20       +1     
  Lines         470      478       +8     
==========================================
+ Hits          457      465       +8     
  Misses         13       13              

Impacted Files	Coverage Δ
src/evaluation/evaluate.jl	93.75% <ø> (ø)
src/metric/base.jl	33.33% <0.00%> (ø)
src/base_recommender.jl	100.00% <100.00%> (ø)
src/baseline/co_occurrence.jl	100.00% <100.00%> (ø)
src/baseline/item_mean.jl	100.00% <100.00%> (ø)
src/baseline/most_popular.jl	100.00% <100.00%> (ø)
src/baseline/threshold_percentage.jl	100.00% <100.00%> (ø)
src/baseline/user_mean.jl	100.00% <100.00%> (ø)
src/data_accessor.jl	100.00% <100.00%> (ø)
src/evaluation/cross_validation.jl	100.00% <100.00%> (ø)
... and 10 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 551019e...2d590fe. Read the comment docs.