Define mapfoldl/foldl for static arrays

[tkf]

This PR includes some improvements discussed in #702 (comment):

• Document how _InitialValue works and explain some idea behind it.
• Use _InitialValue instead of passing around kw.data :: NamedTuple.
• Use _InitialValue to collapse the foldl implementations with and without init into one.
• Add mapfoldl and foldl on static arrays:
  1. Rename the reduce "kernels" _mapreduce to _mapfoldl.
  2. Add _mapreduce back as an alias to _mapfoldl.
     • This way, functions (e.g., sum) wrapping reduce can just call _mapreduce.
     • I added this indirection to avoid another big diff in the future RFC: Add divide-and-conquer _mapreduce #751. It may make sense to use the divide-and-conquer version for _mapreduce Discussion: should we use divide-and-conquer reduce for instruction-level parallelism? #748.
  3. Define Base.mapfoldl and Base.foldl as a simple wrapper of _mapfoldl.

It's not a tiny diff but mostly all the changes are simple refactoring.

[coveralls]

Coverage decreased (-0.06%) to 82.397% when pulling cf9a190 on tkf:foldl into c808bdd on JuliaArrays:master.

[tkf]

Base.reduce_first wasn't called before so sum(SA[true]) returned true; now it returns 1 :: Int. This is consistent with sum([true]).

[c42f]

Looks great to me!